{
  "version": 3,
  "sources": ["ssg:https://framerusercontent.com/modules/YVgynkVt96a0ES9Gk8yz/VEx2WPNLcsYT7stAmBDF/eo4RAmtig-10.js"],
  "sourcesContent": ["import{jsx as e,jsxs as t}from\"react/jsx-runtime\";import{Link as a}from\"framer\";import{motion as n}from\"framer-motion\";import*as i from\"react\";export const richText=/*#__PURE__*/t(i.Fragment,{children:[/*#__PURE__*/e(\"h5\",{children:/*#__PURE__*/e(\"br\",{className:\"trailing-break\"})}),/*#__PURE__*/t(\"h5\",{children:[\"Get an \",/*#__PURE__*/e(a,{href:\"https://toloka.ai/rlhf-guide?utm_source=blog_article&utm_medium=organic&utm_campaign=rlhf_guide&utm_content=banner\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!0,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:\"in-depth guide\"})}),\" to RLHF for superior results. Make models safer and more accurate by using expert data.\"]}),/*#__PURE__*/e(a,{href:\"https://toloka.ai/large-language-models/reinforcement-learning/?utm_source=blog&utm_medium=banner&utm_campaign=rlhf&utm_content=llmalignment\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!0,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{className:\"framer-image\",\"data-preset-tag\":\"img\",children:/*#__PURE__*/e(\"img\",{alt:\"\",className:\"framer-image\",height:\"240\",src:\"https://framerusercontent.com/images/hirqqayfeDZ30QBHLtATp1Fjrpw.png\",srcSet:\"https://framerusercontent.com/images/hirqqayfeDZ30QBHLtATp1Fjrpw.png?scale-down-to=512 512w,https://framerusercontent.com/images/hirqqayfeDZ30QBHLtATp1Fjrpw.png?scale-down-to=1024 1024w,https://framerusercontent.com/images/hirqqayfeDZ30QBHLtATp1Fjrpw.png 1498w\",style:{aspectRatio:\"1498 / 480\"},width:\"749\"})})}),/*#__PURE__*/e(\"p\",{children:\"Large language models (LLMs) capable of generating human-like text are becoming increasingly sophisticated. They have influenced various domains, from customer service and content creation to research and decision-making. Despite their advancements, their primary goal remains the same: to serve human beings in the best possible way.\"}),/*#__PURE__*/e(\"p\",{children:\"To do so, they have to recognize that human values are diverse, complex, and even contradictory sometimes. AI systems become more powerful, yet they should remain controllable by humans. This means that the goals of any LLM should align with human objectives. The AI should pursue goals intended by its human designers without deviating in harmful ways.\"}),/*#__PURE__*/e(\"p\",{children:\"AI or LLM alignment process involves multiple stages and techniques designed to ensure that these models generate outputs consistent with human values, goals, and intentions. This article will explore the nuances of LLM alignment, examining why it is crucial, how it can be achieved, and what it means for the future of AI.\"}),/*#__PURE__*/e(\"h2\",{children:\"What is LLM alignment?\"}),/*#__PURE__*/e(\"p\",{children:'\"A robot may not injure a human being or, through inaction, allow a human being to come to harm.\" This fundamental rule, originally conceived by novelist Isaac Asimov in 1942 in his short story \"Runaround,\" has moved beyond science fiction to influence the real-world development of AI and robotics.'}),/*#__PURE__*/e(\"p\",{children:\"Asimov's First Law of Robotics, as it came to be known, envisioned a world where intelligent machines are bound by ethical guidelines to protect human life. Today, this concept plays a crucial role in shaping how we train our robot assistants and AI systems, ensuring they align with and serve human values and objectives.\"}),/*#__PURE__*/e(\"p\",{children:\"With the advent of generative AI, which amazes the world with its capabilities, particularly in natural language processing, there is a need to exert more control over large language models (LLMs). We increasingly become more reliant on them, which implies more risks. That is why alignment with human values appeared. It prevents large language models from generating harmful or unethical content.\"}),/*#__PURE__*/e(\"p\",{children:\"Basically, AI or LLM alignment provides more control over AI systems. In theory, aligning large language models shouldn't be hard. All that needs to be done is establishing some rules based on human values and then training the model to follow those rules.\"}),/*#__PURE__*/e(\"p\",{children:\"The reality is more complicated because people's goals change depending on their environment. AI alignment refers to ensuring that AI systems behave in a way that respects human values, such as fairness, safety, and rights.\"}),/*#__PURE__*/e(\"h2\",{children:\"Why do you need alignment of large language models?\\xa0\"}),/*#__PURE__*/e(\"p\",{children:\"The answer is that large language models can give harmful responses. Since LLM's main task is to predict the most probable sequence of words, the system can give the sequence it is asked for. As they are trained on vast datasets that include a wide range of human-written text, they may inadvertently generate outputs that reflect unethical, unsafe, or biased information found in that data.\"}),/*#__PURE__*/e(\"p\",{children:\"LLM alignment to human values and goals\"}),/*#__PURE__*/e(\"p\",{children:\"This potential for LLMs to produce toxic, biased, or inaccurate responses highlights the critical need for alignment. Large language models must be harmless, honest, and helpful to be effective and trustworthy. These alignment criteria are essential for a language model.\"}),/*#__PURE__*/e(\"p\",{children:\"Alignment ensures that LLMs consistently meet these criteria by guiding them to generate outputs that are safe, ethical, and aligned with human values. This is essential for both responsible deployment of these models, fostering trust, and ensuring they serve as positive societal tools.\"}),/*#__PURE__*/e(\"p\",{children:\"LLMs operate in various contexts, interacting with users with different expectations, needs, and cultural backgrounds. Alignment plays a key role in ensuring these models respond appropriately in different situations, respecting cultural norms and individual differences. This is especially important in applications distributed globally, where the same model may be used by millions of people with very different perspectives.\"}),/*#__PURE__*/e(\"h2\",{children:\"Methods of LLM alignment\"}),/*#__PURE__*/e(\"h3\",{children:\"Reinforcement Learning from Human Feedback (RLHF)\"}),/*#__PURE__*/t(\"h5\",{children:[/*#__PURE__*/e(a,{href:\"https://toloka.ai/rlhf-guide?utm_source=blog_article&utm_medium=organic&utm_campaign=rlhf_guide&utm_content=banner\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!0,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:\"Get a comprehensive guide\"})}),\" for superior RLHF. Train safer, more accurate models with expert data.\"]}),/*#__PURE__*/e(\"p\",{children:\"A large language model needs support to understand what to generate. For this purpose, the Reinforcement Learning from Human Feedback (RLHF) method was developed, where a person gives human feedback to the model.\"}),/*#__PURE__*/e(\"p\",{children:\"Incorporating direct input from human evaluators into the training process helps refine the model's behavior so that its outputs better align with what humans consider helpful, ethical, and appropriate. Here\u2019s how RLHF works.\"}),/*#__PURE__*/e(\"h3\",{children:\"Pretraining on large datasets\"}),/*#__PURE__*/e(\"p\",{children:\"An LLM is first trained on large datasets using standard supervised learning techniques. This training enables the model to generate coherent and contextually relevant text based on patterns learned from the data. However, since the data often contains biased or harmful content, the model\u2019s outputs may reflect those issues.\"}),/*#__PURE__*/e(\"h3\",{children:\"Human feedback collection\"}),/*#__PURE__*/e(\"p\",{children:\"After the initial training, a model is fine-tuned using human feedback. Human evaluators are asked to rank different outputs generated by the model for the same input prompt. For example, they might rank responses based on how helpful, truthful, or non-toxic they are.\"}),/*#__PURE__*/e(\"h3\",{children:\"Training a reward model\"}),/*#__PURE__*/e(\"p\",{children:\"The feedback from human evaluators is used to train a reward model. Based on the rankings provided by evaluators, this model predicts how well a given output aligns with human preferences. The reward model assigns a score to each output, reflecting its alignment with the desired values.\"}),/*#__PURE__*/e(\"h3\",{children:\"Reinforcement learning\"}),/*#__PURE__*/e(\"p\",{children:\"A LLM is then fine-tuned using reinforcement learning: it is trained to maximize the rewards predicted by the reward model. A language model outputs a response according to its current strategy or policy to create responses. The reward model then assesses the output and gives it a reward. In that way, the model learns to generate outputs that are more likely to be ranked highly by human evaluators. This means that the model aligns its behavior with human expectations.\"}),/*#__PURE__*/e(\"p\",{children:\"In RLHF, once the reward model is trained using human feedback, it generates a reward signal for the outputs produced by the LLM. Proximal Policy Optimization (PPO) is often used as the reinforcement learning algorithm to optimize the LLM's policy (its strategy for generating outputs) based on the reward signals the reward model provides. The original policy gets modified in a way that maximizes the chances of LLM to deliver better and higher-ranked results in the future.\"}),/*#__PURE__*/e(\"h2\",{children:\"Proximal Policy Optimization (PPO)\"}),/*#__PURE__*/e(\"p\",{children:\"PPO performs updates to the LLM\u2019s policy in a way that maximizes the expected reward. This involves adjusting the model's parameters so that it produces outputs that are more likely to receive high scores from the reward model, which reflects human preferences.\"}),/*#__PURE__*/e(\"p\",{children:\"PPO is a policy gradient method that introduces specific innovations to address some of the challenges associated with standard policy gradient approaches. Unlike traditional policy gradient methods, which update the policy with a single step, PPO performs multiple epochs of updates using the same batch of data.\"}),/*#__PURE__*/e(\"p\",{children:\"The typical objective in policy gradient methods is to maximize the expected reward. Policy gradient methods directly optimize the policy (a mapping from states of the environment to actions to be taken when in those states) by adjusting the parameters of a policy model (usually a neural network) to maximize the expected cumulative reward. The core idea is to compute the gradient of the expected reward with respect to the policy parameters and use this gradient to perform gradient ascent, thereby improving the policy over time.\"}),/*#__PURE__*/e(\"p\",{children:\"In traditional policy gradient methods like Policy Gradient (PG) or REINFORCE, the updates to the policy can be large, potentially leading to drastic changes that destabilize learning. PPO was developed to address the issues of instability and inefficiency in traditional policy gradient methods.\"}),/*#__PURE__*/e(\"p\",{children:\"In PPO and other RL algorithms, the policy is typically represented by a neural network as a parameterized function that maps the environment's current states to possible actions. The network is parameterized by weights and biases, which are adjusted during training to optimize the policy.\"}),/*#__PURE__*/e(\"p\",{children:\"The key innovation in PPO is using a clipped objective function that constrains the policy updates, preventing them from becoming too large and destabilizing the learning process. PPO\u2019s objective function is called the clipped surrogate loss function. The surrogate objective function examines the ratio of the probability of action given the current policy to the probability given the reference policy multiplied by an advantage function. The advantage function estimates if an action is better than the average action in a given state.\"}),/*#__PURE__*/e(\"p\",{children:\"Surrogate loss in PPO is determined based on the ratio of the probability of acting on the current policy to the probability of performing the same action under the reference policy. This ratio is then utilized to readjust the policy toward actions that come with a greater reward while keeping the updates from being too extreme. A clipping mechanism is implemented to constrain the amount of those updates, thus keeping the training process stable.\"}),/*#__PURE__*/e(\"p\",{children:\"PPO modifies the standard policy gradient objective to ensure that updates do not deviate too much from the current policy. The clipping mechanism ensures that the update does not push the policy too far in a single step, maintaining stability.\"}),/*#__PURE__*/e(\"p\",{children:\"The surrogate loss function in PPO is central to the algorithm's ability to update policies in a stable and controlled manner. By clipping the probability ratio between the new and old policies, PPO ensures that updates are neither too aggressive nor too conservative, striking a balance that enables efficient and reliable learning.\"}),/*#__PURE__*/e(\"h3\",{children:\"Direct Preference Optimization (DPO)\"}),/*#__PURE__*/t(\"p\",{children:[\"Direct Preference Optimization (DPO), introduced in the 2023 paper \",/*#__PURE__*/e(a,{href:\"https://arxiv.org/abs/2305.18290\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:'\"Your Language Model is Secretly a Reward Model\"'})}),\" by Rafailov et al. from Stanford, represents a significant advancement in aligning large language models with human preferences. Unlike traditional methods that might involve complex processes like reward modeling in reinforcement learning from human feedback, DPO simplifies the pipeline by directly adjusting model outputs to align with human preferences.\"]}),/*#__PURE__*/e(\"p\",{children:\"One of the first steps in RLHF is to train a reward model based on human preference data and fine-tune the language model according to these preferences using reinforcement learning (RL) algorithms like PPO. Methods similar to RLHF have traditionally been used to help guide large language models (LLMs), as the unsupervised nature of their learning still makes them difficult to control accurately. Despite its effectiveness, the RLHF poses challenges of complexity and stability, especially in tailoring the reward model and training policies that optimize that reward.\"}),/*#__PURE__*/e(\"p\",{children:\"One of DPO's key advantages is its simplicity. Unlike RLHF, DPO skips a reward model step and eliminates the need for traditional reinforcement learning, directly using feedback to adjust the model's behavior. It uses advanced mathematics to prove that instead of needing a separate reward model, the model itself can figure out what\u2019s right and wrong as it learns.\"}),/*#__PURE__*/e(\"p\",{children:\"DPO shows that policies and rewards can be combined into a single training step instead of multiple stages. It also changes the way we think about the goal of RLHF: instead of treating rewards and policies separately, it treats rewards as something that can be directly figured out from how likely the AI is to make certain decisions.\"}),/*#__PURE__*/e(\"p\",{children:\"A reward in DPO can be defined as a function of policy probabilities or a function of the probabilities of the model\u2019s outputs. The policy assigns probabilities to different actions (model\u2019s outputs) for any given situation. For instance, if the policy is 80% likely to choose action A and 20% likely to choose action B in a specific situation, these probabilities represent how the policy decides among the available actions.\"}),/*#__PURE__*/e(\"p\",{children:'When we say that a reward is a \"function of policy probabilities,\" we mean that the reward or the outcome can be determined based on these probabilities. This means the reward signal is implicitly built into the probabilities the LLM assigns to different responses. Instead of defining rewards in isolation, they are now expressed directly in terms of how likely the AI is to take certain actions.'}),/*#__PURE__*/e(\"p\",{children:\"DPO uses a mathematical approach to show that you can directly optimize the model\u2019s performance by focusing on probabilities of outputs (how likely the model is to produce certain responses). This eliminates the need for a separate reward model. Here, LLM itself acts as a reward model. DPO leverages preference data, where pairs of actions (or sequences of actions) are compared, and one is labeled as preferable over the other.\"}),/*#__PURE__*/e(\"p\",{children:'When training, the loss function used in DPO is designed to encourage the LLM to increase the probability of responses that are preferred by humans (better completions) and decrease the probability of responses that are less preferred (worse completions). The LLM effectively learns to \"self-assess\" its outputs based on how likely it is to generate preferred responses. By optimizing its own output probabilities, the LLM effectively acts as its own reward model.'}),/*#__PURE__*/e(\"p\",{children:\"DPO treats the policy optimization task as a binary classification problem, employing a binary cross-entropy objective. In other words, it chooses the best of two answers based on the preference data, thereby directly updating the model's policy to improve its results. A binary cross-entropy loss measures how well the model's predictions align with the provided preference labels, i.e. it compares the responses generated by language models. Minimizing this loss directly updates the model's policy.\"}),/*#__PURE__*/e(\"h3\",{children:\"Kahneman-Tversky Optimization (KTO)\"}),/*#__PURE__*/e(\"p\",{children:\"Kahneman-Tversky Optimization (KTO) loss function represents a new approach to training language models. It focuses on maximizing the utility of the generated outputs rather than just improving the log-likelihood of preferences, as is commonly done in traditional methods.\"}),/*#__PURE__*/t(\"p\",{children:[\"KTO is a human-centered approach to training language models. It is named after Daniel Kahneman and Amos Tversky, who developed Prospect Theory, first presented in a paper titled \",/*#__PURE__*/e(a,{href:\"https://web.mit.edu/curhan/www/docs/Articles/15341_Readings/Behavioral_Decision_Theory/Kahneman_Tversky_1979_Prospect_theory.pdf\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:\"Prospect Theory: An Analysis of Decision under Ris\"})}),\"k. This theory is well-known for its insights into how people make decisions under uncertainty and evaluate potential gains and losses.\"]}),/*#__PURE__*/e(\"h4\",{children:\"Prospect theory\"}),/*#__PURE__*/t(\"p\",{children:[\"To understand KTO better, we first have to delve deeper into the Prospect Theory suggested by \",/*#__PURE__*/e(a,{href:\"https://www.worldscientific.com/doi/abs/10.1142/9789814417358_0006\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:\"Kahneman and\\xa0Tversky\"})}),\". Prospect Theory is a way to understand how people make decisions when uncertain about the outcomes, especially when choosing between options involving risks, like winning or losing money. Prospect theory says that losing something hurts more than gaining the same thing feels good. In other words, people hate losing more than they like winning.\"]}),/*#__PURE__*/e(\"p\",{children:'People don\u2019t think of money or other outcomes in terms of absolute wealth or utility. Instead, they compare what they have now (their \"reference point\") to what they could have. The reference point is typically the current state, but it can be influenced by expectations, norms, or recent experiences.'}),/*#__PURE__*/e(\"p\",{children:\"Prospect theory suggests that the value function is steeper for losses than gains, indicating that losses are felt more intensely than equivalent gains. This means that the pain of losing is psychologically about twice as powerful as the pleasure of gaining. For example, if you expect to get $200 and only get $100, you might feel disappointed, even though you\u2019re still $100 richer.\"}),/*#__PURE__*/e(\"p\",{children:\"According to the theory, people do not weigh probabilities linearly. We tend to overweight small probabilities and underweight large probabilities. This means that unlikely events are perceived as more likely than they are, while very likely events are perceived as less certain. People might overestimate the chance of something unlikely, like winning the lottery, and underestimate the chance of something very likely, like getting rained on if it's cloudy.\"}),/*#__PURE__*/e(\"p\",{children:\"Prospect theory helps explain why people sometimes make choices that seem illogical or against their best interests. It shows that people\u2019s decisions are influenced by how they perceive risks and rewards, not just by the actual numbers. They are influenced by their fear of loss, how they perceive their current situation, and how they understand probabilities.\"}),/*#__PURE__*/e(\"p\",{children:\"The prospect theory of Kahneman and Tversky shows us that humans perceive outcomes, particularly those involving risk, in a biased but predictable way. One well-known aspect is loss aversion\u2014people tend to fear losses more than they value gains of the same size. The theory highlights how these biases affect human decision-making. These biases are implicitly considered when we try to align LLMs with human preferences.\"}),/*#__PURE__*/e(\"h3\",{children:\"Human-Aware Loss Functions (HALOs)\"}),/*#__PURE__*/e(\"p\",{children:'The algorithms used to train LLMs (like DPO and PPO) are called human-aware loss functions (HALOs) because they incorporate human biases and decision-making tendencies. These loss functions are \"human-aware\" because they align the model\u2019s outputs with how humans actually perceive and value different outcomes.'}),/*#__PURE__*/e(\"p\",{children:'Such functions unintentionally incorporate biases of human perception, as prospect theory describes. Despite these existing methods, the utility functions they use (i.e., the mathematical representation of what is considered \"good\" or \"bad\" from a human perspective) do not fully align with those described in prospect theory.'}),/*#__PURE__*/t(\"p\",{children:['In a paper titled \"',/*#__PURE__*/e(a,{href:\"https://arxiv.org/pdf/2402.01306\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:\"KTO: Model Alignment as Prospect Theoretic Optimization\"})}),',\" KTO is introduced as a new HALO that directly uses a utility function based on prospect theory. It directly maximizes the utility of the outputs generated by the LLM rather than attempting to predict human preferences based on likelihood scores.']}),/*#__PURE__*/e(\"p\",{children:\"The success of HALOs over non-HALOs simpler techniques like cross-entropy minimization, which does not account for human biases and focuses on predicting exact words, is partly due to their alignment with human biases.\"}),/*#__PURE__*/e(\"h3\",{children:\"What is KTO?\"}),/*#__PURE__*/e(\"p\",{children:\"Human feedback changes over time. However, traditional approaches for LLM alignment, like DPO, mostly rely on consistent datasets of human feedback. Thus, they lack the ability to capture all the nuances of human perspectives and goals. This is where KTO comes to the rescue, as its roots lie in the psychological ideas of Kahneman and Tversky's perspective theory.\"}),/*#__PURE__*/e(\"p\",{children:\"Conventional alignment approaches typically focus on maximizing the log-likelihood of human preferences, which involves adjusting the model to predict better which outputs are preferred based on past data. Instead of focusing solely on these preferences, KTO aims to maximize the utility of the language model\u2019s outputs. Utility here refers to the value or satisfaction the outputs provide users, aligning more directly with human goals and preferences.\"}),/*#__PURE__*/e(\"p\",{children:\"Unlike traditional methods that require detailed preference data, which is often hard to obtain, KTO only needs a simple binary signal\u2014whether an output is desirable or undesirable. This makes it much easier to implement in real-world scenarios where gathering preference data is difficult and expensive.\"}),/*#__PURE__*/e(\"p\",{children:\"The traditional approach to aligning large language models (LLMs) with human values and preferences involves supervised fine-tuning, after which the model is further refined using methods like reinforcement learning from human feedback (RLHF) or direct preference optimization (DPO). These methods require paired preference data, where there are examples of multiple outputs for each input, and humans have labeled one output as better than the others. This helps the model learn which types of responses humans prefer.\"}),/*#__PURE__*/e(\"p\",{children:\"The problem with this traditional pipeline is that collecting paired preference data is difficult. It requires a lot of effort to gather examples where humans have compared different outputs and selected the best one. This data collection process is time-consuming and expensive.\"}),/*#__PURE__*/e(\"p\",{children:\"KTO simplifies this process by requiring only a simpler form of feedback: whether a given output is desirable or undesirable for a specific input. Instead of detailed comparisons between different outputs, KTO needs a yes/no signal indicating whether the output is good or bad.\"}),/*#__PURE__*/t(\"p\",{children:[\"According to the \",/*#__PURE__*/e(a,{href:\"https://arxiv.org/html/2402.01306v1\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:\"paper KTO: Model Alignment as Prospect Theoretic Optimization\"})}),\", KTO performs better than current preference-based methods across models of different sizes (from 1 billion to 30 billion parameters).\"]}),/*#__PURE__*/e(\"h3\",{children:\"Stages of KTO\"}),/*#__PURE__*/e(\"p\",{children:\"Here\u2019s how KTO works:\"}),/*#__PURE__*/t(\"p\",{children:[/*#__PURE__*/e(\"strong\",{children:\"Outputs generation. \"}),\"The language model, during training, generates outputs (like sentences or paragraphs) based on given inputs. These outputs are evaluated as whole pieces of text, not just individual words, which allows the model to focus on producing meaningful and contextually appropriate content.\"]}),/*#__PURE__*/t(\"p\",{children:[/*#__PURE__*/e(\"strong\",{children:\"Outputs evaluation. \"}),\"Each generated output is then assessed using the utility function. This evaluation determines how well the output meets the desired criteria. The utility function is based on Kahneman-Tversky\u2019s prospect theory. The function can consider various factors like how relevant, coherent, or appropriate an output is according to specific criteria. The output receives a utility score that indicates its desirability\u2014essentially, how much a human would value that output.\"]}),/*#__PURE__*/t(\"p\",{children:[/*#__PURE__*/e(\"strong\",{children:\"Model optimization.\"}),\" The model's internal parameters determine how it generates text and are adjusted based on the utility scores. This process aims to increase the chances that the model will produce outputs with higher utility scores in the future, meaning outputs that are more aligned with what humans want.\"]}),/*#__PURE__*/t(\"p\",{children:[/*#__PURE__*/e(\"strong\",{children:\"Iterative process of training.\"}),\" This is a continuous loop where the model generates outputs, receives feedback via the utility scores, and updates its parameters accordingly. This iterative process teaches the model to consistently produce outputs that are better aligned with the utility function\u2019s criteria for desirability.\"]}),/*#__PURE__*/e(\"p\",{children:\"In KTO, instead of just trying to predict the next word or fit to pre-labeled data, the language model is trained to generate outputs that maximize a utility score based on human preferences. This utility-driven approach is extremely useful for tasks where the output quality is subjective or where specific output traits are highly valued. By focusing directly on what makes an output desirable, KTO helps create language models that are better aligned with human needs and values.\"}),/*#__PURE__*/e(\"h3\",{children:\"Advantages of KTO Over DPO and PPO\"}),/*#__PURE__*/e(\"h4\",{children:\"Data requirements\"}),/*#__PURE__*/e(\"p\",{children:\"PPO generally requires a well-defined reward model, and DPO relies on paired preference data, which is harder to gather and requires more nuanced judgments. KTO requires only a binary signal indicating whether an output is desirable or undesirable. This is much easier to collect than detailed preference data.\"}),/*#__PURE__*/e(\"h4\",{children:\"Direct utility maximization\"}),/*#__PURE__*/e(\"p\",{children:\"DPO does not explicitly maximize a utility function; instead, it increases the likelihood of preferred outputs based on collected preferences. PPO optimizes a reward signal through reinforcement learning, which can be indirectly aligned with utility but often requires careful tuning and may not reflect human biases as directly as KTO.\"}),/*#__PURE__*/e(\"p\",{children:\"KTO focuses on directly maximizing the utility of model outputs based on human-like evaluation criteria derived from prospect theory rather than just matching preferences. This leads to outputs that are more aligned with human values.\"}),/*#__PURE__*/e(\"h4\",{children:\"Real-world applicability\"}),/*#__PURE__*/e(\"p\",{children:\"The reduced need for specific preference data and its reliance on more abundant, simple feedback signals make KTO more practical and easier to implement in real-world scenarios.\"}),/*#__PURE__*/e(\"h2\",{children:\"PPO vs. DPO vs. KTO\"}),/*#__PURE__*/e(\"img\",{alt:\"\",className:\"framer-image\",height:\"504\",src:\"https://framerusercontent.com/images/iGdiiDEtnMHYAS5lUea2BcYQToQ.png\",srcSet:\"https://framerusercontent.com/images/iGdiiDEtnMHYAS5lUea2BcYQToQ.png?scale-down-to=512 512w,https://framerusercontent.com/images/iGdiiDEtnMHYAS5lUea2BcYQToQ.png?scale-down-to=1024 1024w,https://framerusercontent.com/images/iGdiiDEtnMHYAS5lUea2BcYQToQ.png 1162w\",style:{aspectRatio:\"1162 / 1008\"},width:\"581\"}),/*#__PURE__*/e(\"h2\",{children:\"RLAIF\"}),/*#__PURE__*/e(\"p\",{children:\"Reinforcement Learning with AI Feedback (RLAIF) is an innovative approach to aligning large language models (LLMs) with desired behaviors and outcomes using feedback generated by artificial intelligence rather than human input alone. This method enhances traditional alignment methods, such as reinforcement learning with human feedback and aims to overcome some of the limitations of human-based feedback systems.\"}),/*#__PURE__*/e(\"p\",{children:\"The primary goal of RLAIF is to reduce reliance on costly and labor-intensive human feedback by employing a readymade LLM as a preference labeler. A significant advantage of RLAIF is its ability to drastically reduce human annotations' costs.\"}),/*#__PURE__*/e(\"p\",{children:\"Instead of fine-tuning language models with human feedback using human evaluators (as in RLHF), RLAIF leverages AI systems to generate this feedback. Such AI systems are typically trained to simulate human preferences and judgments, creating a feedback loop that can be more scalable and consistent than relying solely on human input.\"}),/*#__PURE__*/e(\"p\",{children:\"However, developing AI models capable of accurately simulating human preferences is a complex task. These models must be carefully designed and trained to reflect diverse and nuanced human judgments. Despite the advantages of AI-generated feedback, human oversight is still crucial to ensure that the model remains aligned with human values and ethical standards. RLAIF may require periodic human validation to ensure the AI feedback is steering the model in the right direction.\"}),/*#__PURE__*/e(\"h3\",{children:\"How RLAIF fits into the LLM alignment process\"}),/*#__PURE__*/e(\"p\",{children:\"RLAIF can be used alongside traditional human feedback methods to create a more robust and scalable alignment process. AI feedback can handle most of the LLM alignment work, with human feedback providing additional validation and fine-tuning. By incorporating artificial intelligence for feedback, RLAIF can accelerate the alignment process, which offers faster iterations and improving LLM performance.\"}),/*#__PURE__*/t(\"p\",{children:[\"In the paper \",/*#__PURE__*/e(a,{href:\"https://arxiv.org/abs/2212.08073\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:/*#__PURE__*/e(\"strong\",{children:'\"Constitutional AI: Harmlessness from AI Feedback\"'})})}),\" by Bai et al., the authors explore how Reinforcement Learning from AI Feedback (RLAIF) can be used to scale the alignment process for large language models, focusing on the benefits of using AI-generated feedback instead of traditional human feedback. They suggest that integrating RLAIF with RLHF could harness the advantages of both methods, providing a more robust approach to LLM alignment and training.\"]}),/*#__PURE__*/e(\"h3\",{children:\"How does RLAIF work?\"}),/*#__PURE__*/e(\"h4\",{children:\"Model pre-training\"}),/*#__PURE__*/e(\"p\",{children:\"The RLAIF process starts with a model already pre-trained on a broad corpus of text. This model has learned general language patterns and serves as the foundation for further fine-tuning using RLAIF. This feedback model, often an LLM itself, is guided by a set of rules designed to ensure it produces feedback that promotes safety, helpfulness, and honesty.\"}),/*#__PURE__*/e(\"h4\",{children:\"Generating outputs\"}),/*#__PURE__*/e(\"p\",{children:\"Another LLM called the response model intended for training is then fed with various input prompts related to the target task (e.g., summarization, dialogue generation). It generates multiple responses or outputs for each input prompt. These outputs could be different summaries or dialogue responses. These responses are then evaluated by the feedback model. The feedback model assesses these responses and provides numerical preference scores for each response.\"}),/*#__PURE__*/e(\"h4\",{children:\"AI feedback generation\"}),/*#__PURE__*/e(\"p\",{children:\"As mentioned above, instead of using human feedback, RLAIF leverages another LLM, the feedback model, to evaluate the quality of these outputs. Such an AI labeler assigns a reward score to each output, which reflects how well it aligns with predefined criteria like relevance, coherence, or helpfulness.\"}),/*#__PURE__*/e(\"h4\",{children:\"Assigning reward scores\"}),/*#__PURE__*/e(\"p\",{children:\"The feedback model or AI labeler assigns a reward score to each output. These scores reflect how well the outputs align with the desired criteria, effectively quantifying the quality of each response. This process can be enhanced by using techniques like chain-of-thought reasoning, where the labeler provides more detailed and thoughtful evaluations.\"}),/*#__PURE__*/e(\"h4\",{children:\"AI-generated dataset creation\"}),/*#__PURE__*/e(\"p\",{children:\"The feedback model's evaluations produce an AI-generated dataset. This dataset contains prompts, pairs of responses, and corresponding preference scores. This dataset is akin to the human feedback data collected in RLHF, but it\u2019s generated by the AI feedback model.\"}),/*#__PURE__*/e(\"h4\",{children:\"Revision and Critique Phase\"}),/*#__PURE__*/e(\"p\",{children:\"Before moving to fine-tuning, the response model undergoes a revision and critique phase:\"}),/*#__PURE__*/t(\"ul\",{children:[/*#__PURE__*/e(\"li\",{\"data-preset-tag\":\"p\",style:{\"--framer-font-size\":\"12px\",\"--framer-text-color\":\"rgb(0, 0, 0)\",\"--framer-text-decoration\":\"none\"},children:/*#__PURE__*/t(\"p\",{children:[/*#__PURE__*/e(\"strong\",{children:\"Initial revision\"}),\". The response model should be further fine-tuned on a dataset that includes responses revised through an iterative process. So, harmful responses are identified and removed during that stage using a separate helpful RLHF model.\"]})}),/*#__PURE__*/e(\"li\",{\"data-preset-tag\":\"p\",style:{\"--framer-font-size\":\"12px\",\"--framer-text-color\":\"rgb(0, 0, 0)\",\"--framer-text-decoration\":\"none\"},children:/*#__PURE__*/t(\"p\",{children:[/*#__PURE__*/e(\"strong\",{children:\"Critique process\"}),\". During the critique phase, the RLHF model generates responses to various prompts and may identify harmful elements in its own responses. This process involves iteratively refining the responses to ensure they align with safety and helpfulness guidelines.\"]})})]}),/*#__PURE__*/e(\"h4\",{children:\"Supervised Learning\"}),/*#__PURE__*/e(\"p\",{children:\"Following the revision phase, the response model undergoes supervised learning. The model is trained using the revised dataset to ensure it generates outputs that adhere to the constitutional principles - principles for an AI to follow to make itself harmless, helpful, and honest. This step refines the model to produce safer and more aligned responses.\"}),/*#__PURE__*/e(\"h4\",{children:\"Harmlessness dataset generation\"}),/*#__PURE__*/e(\"p\",{children:\"In this stage, the feedback model evaluates responses generated by the response model to potentially harmful queries. This evaluation produces preference scores for each response. The feedback model uses its constitution to ensure these scores align with safety and quality guidelines. The model calculates and normalizes the log probabilities for each response to create a set of preference data, which includes the prompt, possible completions, and their associated probabilities.\"}),/*#__PURE__*/e(\"h4\",{children:\"Preference model training\"}),/*#__PURE__*/e(\"p\",{children:\"The dataset generated from the previous step is used to train the preference model. This model learns to assign preference scores to responses based on the AI-generated feedback data. The trained preference model can now evaluate new responses and provide preference scores.\"}),/*#__PURE__*/e(\"h4\",{children:\"Final reinforcement learning (RL) Stage\"}),/*#__PURE__*/e(\"p\",{children:\"In the final stage, the trained preference model is used to fine-tune the response model through reinforcement learning. The model is adjusted based on the preference scores provided by the preference model to improve its performance. The goal is to produce responses that align better with human values and preferences as modeled by the feedback and preference models.\"}),/*#__PURE__*/e(\"p\",{children:\"Reinforcement Learning from AI Feedback represents a significant advancement in the alignment of large language models and offers a promising alternative to traditional methods that rely heavily on human feedback. They have not yet reached the full extent of their capabilities and still have to operate under human supervision, but in time, they may become more powerful.\"}),/*#__PURE__*/e(\"h2\",{children:\"What are some challenges associated with LLM Alignment?\"}),/*#__PURE__*/e(\"h3\",{children:\"Exploitation of model weaknesses through adversarial attacks\"}),/*#__PURE__*/e(\"p\",{children:\"Adversarial exploitation in the context of large language models (LLMs) refers to the deliberate manipulation of these models to produce harmful, biased, or otherwise undesirable outputs. Adversaries can craft specific inputs designed to trick the LLM into generating outputs that it wouldn\u2019t normally produce. These inputs might exploit subtle weaknesses or biases in the model, leading to outputs that could be offensive, dangerous, or misleading.\"}),/*#__PURE__*/e(\"p\",{children:\"If an LLM has been trained on biased data, adversaries might exploit these biases to generate content that reinforces stereotypes, spreads misinformation, or causes harm. Even if the model is generally aligned, carefully crafted prompts can trigger these inherent biases. Large language models are also context-dependent and can produce vastly different outputs based on subtle changes in input. Adversaries can manipulate the context in which a question or prompt is given to steer the model towards undesirable outcomes.\"}),/*#__PURE__*/e(\"h3\",{children:\"Interpretability and transparency of language models\"}),/*#__PURE__*/e(\"p\",{children:'With the complexity of LLMs increasing, it becomes more challenging to understand how they create specific outputs. The decision-making process of these models is often a \"black box,\" meaning it\u2019s not transparent or easily interpretable by humans. This opacity makes it challenging to identify when, how, or why a model might produce an output that is misaligned with human values.'}),/*#__PURE__*/e(\"p\",{children:\"When complex models generate harmful or incorrect outputs, it can be difficult to trace these errors back to specific aspects of the model's architecture, data, or training process. This makes it hard to diagnose and fix problems, especially in a timely manner. The more complex a model, the harder it is to ensure that it will behave as intended across all possible scenarios. This increases the risk of misalignment, where the model\u2019s behavior deviates from what is considered safe or ethical.\"}),/*#__PURE__*/e(\"h3\",{children:\"Subjectivity and context-dependence of human values\"}),/*#__PURE__*/e(\"p\",{children:\"Human values are often not absolute but rather depend on context. For example, telling a lie might generally be considered wrong, but many people would find it acceptable or even necessary in certain situations, such as to protect someone\u2019s feelings. Encoding such nuanced, context-dependent judgments into an LLM is challenging because the model needs to understand and appropriately respond to various situations.\"}),/*#__PURE__*/e(\"p\",{children:\"Concepts like fairness, justice, and kindness are subjectively interpreted, even within a single society. Different people might have different thresholds for what they consider fair or just, depending on their perspectives and experiences. Aligning an LLM with such subjective values is difficult because there isn\u2019t always a clear, objective standard to follow.\"}),/*#__PURE__*/e(\"p\",{children:\"Social norms, ethical standards, and cultural values can change significantly within a generation. An LLM that is consistent with today's values may become inconsistent with them as they evolve. Keeping AI systems up-to-date with the latest societal norms and values is difficult.\"}),/*#__PURE__*/e(\"h2\",{children:\"A path towards Responsible AI\"}),/*#__PURE__*/e(\"p\",{children:\"Traditional LLM alignment approaches, such as RLHF, provide models with human preference data according to human judgments and lay the foundation for LLM alignment, incorporating human preferences directly into the learning process.\"}),/*#__PURE__*/e(\"p\",{children:\"Building on this, novel techniques such as Direct Preference Optimization (DPO) simplify the process by eliminating the need for a separate reward model and directly fine-tuning the model based on human preferences. Kahneman-Tversky Optimization (KTO) introduces a new perspective by applying principles from behavioral economics to maximize the utility of outputs, aligning them more closely with how humans make decisions.\"}),/*#__PURE__*/e(\"p\",{children:\"Reinforcement Learning from AI Feedback (RLAIF) represents a significant leap forward, using AI-generated feedback to reduce the dependency on costly human annotations, thereby enhancing scalability and efficiency. Unlike RLHF, where human feedback is collected to rank or rate model outputs, RLAIF generates feedback through another AI system. This AI system is trained to mimic human preferences and values, providing feedback on the LLM's outputs.\"}),/*#__PURE__*/e(\"p\",{children:\"These advances are critical to developing AI systems that are more in line with human values. As these LLM alignment techniques progress and become integrated, they will play a key role in creating language models that generate high-quality content and do so in a way that is consistent with human values and ethical considerations. These advances represent a significant step towards creating AI technologies that are more reliable, trustworthy, and relevant to society's needs and expectations.\"})]});export const richText1=/*#__PURE__*/t(i.Fragment,{children:[/*#__PURE__*/e(\"p\",{children:\"AI red teaming is a systematic approach to identifying vulnerabilities and potential failures within AI models. This practice involves stress-testing AI systems by simulating real-world scenarios, such as adversarial attacks and other challenging conditions, to assess and improve their robustness, reliability, and security.\"}),/*#__PURE__*/e(\"p\",{children:'The concept of red teaming originally comes from military practice, where a dedicated group, known as the \"red team,\" simulates attacks on an organization or particular unit to expose its problem areas that may be subject to intrusion. This methodology has been adapted to IT security, where it\u2019s used for testing the safety of digital systems.'}),/*#__PURE__*/e(\"p\",{children:\"Red teaming is distinct from penetration testing (Pentesting), which typically involves targeted attacks to exploit suspected vulnerabilities. While pentesting is scoped to specific weaknesses within a system, red teaming is broader, encompassing various adversarial tactics.\\xa0\"}),/*#__PURE__*/e(\"img\",{alt:\"Comparable dimensions of penetration testing and red teaming\",className:\"framer-image\",height:\"225\",src:\"https://framerusercontent.com/images/95EM0pMbQADQrC1bcgT8EZ4n2pI.png\",style:{aspectRatio:\"480 / 450\"},width:\"240\"}),/*#__PURE__*/t(\"p\",{children:[/*#__PURE__*/e(\"em\",{children:\"Comparable dimensions of penetration testing and red teaming. (Source:\"}),/*#__PURE__*/e(a,{href:\"https://www.researchgate.net/publication/346770064_Red_Teams_-_Pentesters_APTs_or_Neither\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:/*#__PURE__*/e(\"em\",{children:\" Red Teams \u2014 Pentesters, APTs, or Neither\"})})}),/*#__PURE__*/e(\"em\",{children:\")\"})]}),/*#__PURE__*/e(\"p\",{children:\"When applied to artificial intelligence, the same principles allow engineers to uncover and mitigate weaknesses in AI models before malicious actors can exploit them. Besides, red teaming AI systems means dealing with possible biases and ethical concerns, pushing the boundaries of what the AI can handle.\"}),/*#__PURE__*/e(\"p\",{children:\"In this article, we\u2019ll explore the significance of red teaming for AI projects, including large language models, examine red teaming tools and strategies, and demonstrate how Toloka can help make your AI models safer and more reliable.\"}),/*#__PURE__*/e(\"h2\",{children:\"What is red teaming for Gen AI?\"}),/*#__PURE__*/t(\"h5\",{children:[\"Dive deeper into how red teaming can improve the security and performance of your AI models, check out our comprehensive guide on red teaming\",/*#__PURE__*/e(a,{href:\"https://toloka.ai/red-teaming-guide?utm_source=blog_article&utm_medium=banner&utm_campaign=red_teaming \",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!0,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:\" here\"})}),\".\"]}),/*#__PURE__*/e(\"p\",{children:\"Generative AI red teaming is a practice designed to assess and strengthen the resilience of generative AI systems. Unlike traditional cybersecurity red teaming, which focuses on identifying vulnerabilities, or red teaming in traditional machine learning, which involves testing models for specific weaknesses like data poisoning or model tampering,\\xa0 involves evaluating how generative models handle unpredictable and potentially harmful inputs.\"}),/*#__PURE__*/e(\"p\",{children:\"This approach goes beyond probing for security flaws\u2014it involves stress-testing the AI\u2019s decision-making processes to uncover unintended behaviors. The complexity of generative AI systems, with their ability to create original content, requires a nuanced approach, where potential misuse, ethical implications, and misinformation spread become key concerns.\"}),/*#__PURE__*/e(\"p\",{children:\"Red teaming for generative AI involves challenging the model to produce outputs it is supposed to avoid and to reveal biases its developers may not have anticipated. The model can be realigned with reinforced security measures when these issues are identified.\"}),/*#__PURE__*/e(\"img\",{alt:\"A toxic prompt and jailbreaking attempt using the original prompt manipulation. \",className:\"framer-image\",height:\"186\",src:\"https://framerusercontent.com/images/Y41CUAOaiw89Z18nABE8PgDKAHg.png\",style:{aspectRatio:\"498 / 372\"},width:\"249\"}),/*#__PURE__*/t(\"p\",{children:[/*#__PURE__*/e(\"em\",{children:\"A toxic prompt and jailbreaking attempt using the original prompt manipulation. Finally, it bypasses the safety mechanisms. Source:\"}),/*#__PURE__*/e(a,{href:\"https://arxiv.org/pdf/2408.11727\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:/*#__PURE__*/e(\"em\",{children:\" Efficient Detection of Toxic Prompts in Large Language Models\"})})}),/*#__PURE__*/e(\"em\",{children:\")\"})]}),/*#__PURE__*/e(a,{href:\"https://toloka.ai/red-teaming-guide?utm_source=blog&utm_medium=banner&utm_campaign=redteaming&utm_content=airedteamingsafeguardingyouraimodel\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!0,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{className:\"framer-image\",\"data-preset-tag\":\"img\",children:/*#__PURE__*/e(\"img\",{alt:\"\",className:\"framer-image\",height:\"240\",src:\"https://framerusercontent.com/images/nKOyte7LcB3RIZn7InTQun2SJ6Q.png\",srcSet:\"https://framerusercontent.com/images/nKOyte7LcB3RIZn7InTQun2SJ6Q.png?scale-down-to=512 512w,https://framerusercontent.com/images/nKOyte7LcB3RIZn7InTQun2SJ6Q.png?scale-down-to=1024 1024w,https://framerusercontent.com/images/nKOyte7LcB3RIZn7InTQun2SJ6Q.png 1498w\",style:{aspectRatio:\"1498 / 480\"},width:\"749\"})})}),/*#__PURE__*/e(\"h2\",{children:\"Why red teaming is essential for AI models\"}),/*#__PURE__*/e(\"p\",{children:\"AI models must be robust, secure, and fair, particularly those deployed in critical applications. Red teaming is essential because it allows organizations to proactively identify and address potential issues before they manifest in real-world scenarios. This approach helps prevent costly failures, ensures compliance with ethical standards, and generally enhances trust in AI systems.\"}),/*#__PURE__*/e(\"img\",{alt:\"\",className:\"framer-image\",height:\"345\",src:\"https://framerusercontent.com/images/NcdxYIldVmjd8A3p7Q7UHzVXAo.png\",srcSet:\"https://framerusercontent.com/images/NcdxYIldVmjd8A3p7Q7UHzVXAo.png?scale-down-to=512 512w,https://framerusercontent.com/images/NcdxYIldVmjd8A3p7Q7UHzVXAo.png 814w\",style:{aspectRatio:\"814 / 691\"},width:\"407\"}),/*#__PURE__*/t(\"p\",{children:[/*#__PURE__*/e(\"em\",{children:\"In April 2024, New York City Mayor Eric Adams had to respond to criticism over the chatbot intended to help small businesses but was caught giving bizarre and illegal advice. (Source:\"}),/*#__PURE__*/e(a,{href:\"https://www.techdirt.com/2024/04/05/nyc-officials-are-mad-because-journalists-pointed-out-the-citys-new-ai-chatbot-tells-people-to-break-the-law/\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:/*#__PURE__*/e(\"em\",{children:\" Techdirt\"})})}),/*#__PURE__*/e(\"em\",{children:\")\"})]}),/*#__PURE__*/t(\"p\",{children:[\"On 1 August 2024, the\",/*#__PURE__*/e(a,{href:\"https://www.europarl.europa.eu/doceo/document/TA-9-2024-0138_EN.pdf\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:\" EU AI Act\"})}),\" entered into force, with most of its provisions commencing in 2026. It became the first artificial intelligence law in the European Union, obliging businesses to mitigate risks associated with generative AI systems. The 400-page document mentions the concept of red teaming, requiring organizations to carry out and document adversarial testing.\"]}),/*#__PURE__*/e(\"p\",{children:\"In July 2024, the National AI Security Institute released a public draft of its guidance on \u201CManaging Misuse Risk for Dual-Use Foundation Models\u201D in the United States. The draft specifies that red teams should consist of external experts independent of the AI model developer.\"}),/*#__PURE__*/e(\"img\",{alt:\"\",className:\"framer-image\",height:\"100\",src:\"https://framerusercontent.com/images/toi2ie7jJwSTG2buwYlijhRC9wM.png\",srcSet:\"https://framerusercontent.com/images/toi2ie7jJwSTG2buwYlijhRC9wM.png?scale-down-to=512 512w,https://framerusercontent.com/images/toi2ie7jJwSTG2buwYlijhRC9wM.png 961w\",style:{aspectRatio:\"961 / 201\"},width:\"480\"}),/*#__PURE__*/t(\"p\",{children:[/*#__PURE__*/e(\"em\",{children:\"The U.S. AI Safety Institute lists red teaming among the actions essential for measuring the risk of unwanted activities. Source:\"}),/*#__PURE__*/e(a,{href:\"https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.800-1.ipd.pdf\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:/*#__PURE__*/e(\"em\",{children:\" Managing Misuse Risk for Dual-Use Foundation Models\"})})})]}),/*#__PURE__*/t(\"p\",{children:[\"Beyond regulations that specifically mention red teaming, companies must also recognize their broader responsibility for the AI software they use. For example, the US Equal Employment Opportunity Commission\",/*#__PURE__*/e(a,{href:\"https://industrialrelationsnews.ioe-emp.org/industrial-relations-and-labour-law-june-2023/news/article/usa-equal-employment-opportunity-commission-issued-new-guidance-on-the-use-of-artificial-intelligence-in-employment-selection-procedures\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:\" classifies algorithmic decision-making as an employee selection procedure\"})}),\". This means employers are fully accountable for any discrimination arising from AI biases that are difficult to detect and prevent without red teaming.\\xa0\"]}),/*#__PURE__*/e(\"h3\",{children:\"The benefits of implementing red teaming in AI Systems\"}),/*#__PURE__*/e(\"p\",{children:\"Implementing red teaming offers several critical benefits for developing and deploying robust AI systems.\"}),/*#__PURE__*/e(\"h4\",{children:\"Enhanced security and early threat detection\"}),/*#__PURE__*/e(\"p\",{children:\"By simulating adversarial attacks, red teams help identify and mitigate security vulnerabilities. This early detection of potential threats allows organizations to prevent significant damages by addressing risks before they can be exploited.\"}),/*#__PURE__*/e(\"h4\",{children:\"Improved model performance and adaptation to emerging threats\"}),/*#__PURE__*/e(\"p\",{children:\"Stress testing reveals weaknesses that can be addressed to enhance the model's accuracy, reliability, and overall performance. Continuous red teaming also enables organizations to stay ahead of the rapidly evolving landscape of cybersecurity threats.\"}),/*#__PURE__*/e(\"h4\",{children:\"Increased transparency and building stakeholder confidence\"}),/*#__PURE__*/e(\"p\",{children:\"Rigorous generative AI red teaming builds trust among stakeholders by demonstrating a commitment to the AI model's security and fairness. Regularly validated AI systems, especially those assessed by credible external teams, help foster confidence in the system's safety and reliability.\\xa0\"}),/*#__PURE__*/e(\"h4\",{children:\"Regulatory compliance\"}),/*#__PURE__*/e(\"p\",{children:\"As regulations surrounding AI systems become more stringent, AI red teaming ensures that models comply with legal and ethical standards. This approach addresses existing issues but also anticipates and prevents future problems.\"}),/*#__PURE__*/e(\"p\",{children:\"These benefits highlight the value of red teaming as an essential component of AI model development. By allowing engineers to simulate various unwanted scenarios in a controlled environment, red teaming ensures the AI system can appropriately respond to potential threats.\"}),/*#__PURE__*/e(\"h2\",{children:\"AI Red teaming practices\"}),/*#__PURE__*/e(\"p\",{children:\"Red teams are focused on simulating adversarial scenarios and testing the limits of AI models. Here\u2019s a brief overview of some key red-teaming techniques ethical hackers use to enhance the security and reliability of AI systems.\"}),/*#__PURE__*/e(\"h3\",{children:\"Jailbreak prompting\"}),/*#__PURE__*/e(\"p\",{children:\"Jailbreak prompting exposes weaknesses in LLMs by pushing the models to deviate from their safety constraints. This method reveals how models can be manipulated to produce harmful or biased outputs, highlighting potential conflicts between their capabilities and safety protocols.\\xa0\"}),/*#__PURE__*/e(\"p\",{children:\"While cautious, this creative approach can overly restrict the model, leading to frequent evasiveness. Thus, there's a trade-off between making the model helpful by following instructions and keeping it harmless by minimizing the risk of harm.\"}),/*#__PURE__*/e(\"img\",{alt:\"\",className:\"framer-image\",height:\"101\",src:\"https://framerusercontent.com/images/bp0V2HtOL2DcKVBZITeF806FPw.png\",style:{aspectRatio:\"512 / 203\"},width:\"256\"}),/*#__PURE__*/t(\"p\",{children:[/*#__PURE__*/e(\"em\",{children:\"Instructing the model to respond in code instead of natural language can also reveal its learned biases. (Source:\"}),/*#__PURE__*/e(a,{href:\"https://huggingface.co/blog/red-teaming\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:/*#__PURE__*/e(\"em\",{children:\" Hugging Face\"})})}),/*#__PURE__*/e(\"em\",{children:\")\"})]}),/*#__PURE__*/e(\"h3\",{children:\"Human-guided and automated red teaming\"}),/*#__PURE__*/e(\"p\",{children:\"Human intuition and creativity help identify vulnerabilities in an AI system. A red team and its ethical hackers use their expertise to craft inputs that challenge the model's responses and test how well the AI adheres to ethical standards under pressure.\"}),/*#__PURE__*/e(\"p\",{children:\"Also there are software tools used to mimic a wide range of real-world cyberattacks provide a scalable and efficient approach to generative AI red teaming with automated frameworks, which can conduct an unlimited number of attacks on the target system.\"}),/*#__PURE__*/e(\"p\",{children:\"For example, multi-round automatic red-teaming (MART) involves an adversarial LLM and a target LLM working in cycles, where the adversarial LLM generates challenging prompts that the target LLM must learn to handle safely. Another technique, deep adversarial automated red-teaming (DART), dynamically adjusts attack strategies across iterations, further enhancing the model's safety.\"}),/*#__PURE__*/e(\"img\",{alt:\"\",className:\"framer-image\",height:\"277\",src:\"https://framerusercontent.com/images/XGLdCuaHX4OrBM16yPsf7rMlD1g.png\",srcSet:\"https://framerusercontent.com/images/XGLdCuaHX4OrBM16yPsf7rMlD1g.png 439w\",style:{aspectRatio:\"439 / 555\"},width:\"219\"}),/*#__PURE__*/t(\"p\",{children:[/*#__PURE__*/e(\"em\",{children:\"Test cases can be automatically generated by a language model (LM) and replied by the target LM, with failing test cases being found by a classifier. (Source:\"}),/*#__PURE__*/e(a,{href:\"https://arxiv.org/abs/2202.03286\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:/*#__PURE__*/e(\"em\",{children:\" Red Teaming Language Models with Language Models\"})})}),/*#__PURE__*/e(\"em\",{children:\")\"})]}),/*#__PURE__*/e(\"p\",{children:\"Automated tools often rely on pre-existing classifiers to detect undesirable outputs, limiting the adaptability of red teaming to specific models. However, certain approaches are specifically designed to eliminate this problem.\"}),/*#__PURE__*/e(\"img\",{alt:\"\",className:\"framer-image\",height:\"170\",src:\"https://framerusercontent.com/images/xoQMqBdiJlCgwEHRaCJZTnX8fhY.png\",srcSet:\"https://framerusercontent.com/images/xoQMqBdiJlCgwEHRaCJZTnX8fhY.png?scale-down-to=512 512w,https://framerusercontent.com/images/xoQMqBdiJlCgwEHRaCJZTnX8fhY.png 811w\",style:{aspectRatio:\"811 / 340\"},width:\"405\"}),/*#__PURE__*/t(\"p\",{children:[/*#__PURE__*/e(\"em\",{children:\"A framework suggested by MIT scientists in 2023. Their approach presumes starting with sampling from the target model. (Source:\"}),/*#__PURE__*/e(a,{href:\"https://arxiv.org/pdf/2306.09442\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:/*#__PURE__*/e(\"em\",{children:\" Explore, Establish, Exploit: Red-Teaming Language Models from Scratch\"})})}),/*#__PURE__*/e(\"em\",{children:\")\"})]}),/*#__PURE__*/e(\"h3\",{children:\"Tools for an AI red teaming\"}),/*#__PURE__*/e(\"p\",{children:\"A variety of tools have been developed to assist in the process, including specialized datasets, automated frameworks, and evaluation platforms that enhance the scope of red team activities.\\xa0\"}),/*#__PURE__*/t(\"p\",{children:[\"For instance,\",/*#__PURE__*/e(a,{href:\"https://huggingface.co/datasets/ibm/AttaQ\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:\" AttaQ\"})}),\" is an AI red teaming dataset, which includes 1,402 adversarial questions to evaluate LLMs. This dataset serves as a benchmark for measuring potential risks. By utilizing this dataset, researchers can systematically test AI systems, identify ethical concerns, and work towards reducing harmful outcomes.\"]}),/*#__PURE__*/e(\"img\",{alt:\"\",className:\"framer-image\",height:\"252\",src:\"https://framerusercontent.com/images/CLBfMXKhFhQYneewcLeTZttiE.png\",srcSet:\"https://framerusercontent.com/images/CLBfMXKhFhQYneewcLeTZttiE.png?scale-down-to=512 512w,https://framerusercontent.com/images/CLBfMXKhFhQYneewcLeTZttiE.png 1000w\",style:{aspectRatio:\"1000 / 505\"},width:\"500\"}),/*#__PURE__*/t(\"p\",{children:[/*#__PURE__*/e(\"em\",{children:\"In the AttaQ dataset, all questions are divided into seven classes: deception, discrimination, harmful information, substance abuse, sexual content, personally identifiable information, and violence. Source:\"}),/*#__PURE__*/e(a,{href:\"https://huggingface.co/datasets/ibm/AttaQ\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:/*#__PURE__*/e(\"em\",{children:\" HuggingFace\"})})}),/*#__PURE__*/e(\"em\",{children:\")\"})]}),/*#__PURE__*/e(\"p\",{children:\"The open-source evaluation framework HarmBench identifies several desirable properties previously overlooked in red teaming evaluations, providing a systematic approach to benchmarking.\"}),/*#__PURE__*/e(\"img\",{alt:\"\",className:\"framer-image\",height:\"390\",src:\"https://framerusercontent.com/images/bBRz7hK5rSojAZNVIxa53BPZQ.png\",srcSet:\"https://framerusercontent.com/images/bBRz7hK5rSojAZNVIxa53BPZQ.png?scale-down-to=512 512w,https://framerusercontent.com/images/bBRz7hK5rSojAZNVIxa53BPZQ.png?scale-down-to=1024 1024w,https://framerusercontent.com/images/bBRz7hK5rSojAZNVIxa53BPZQ.png 1494w\",style:{aspectRatio:\"1494 / 781\"},width:\"747\"}),/*#__PURE__*/t(\"p\",{children:[/*#__PURE__*/e(\"em\",{children:\"Illustration of the standardized evaluation pipeline, given an attack method and a model. A diverse set of behaviors is transformed into test cases. Source:\"}),/*#__PURE__*/e(a,{href:\"https://arxiv.org/pdf/2402.04249\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:/*#__PURE__*/e(\"em\",{children:\" HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal\"})})}),/*#__PURE__*/e(\"em\",{children:\")\"})]}),/*#__PURE__*/t(\"p\",{children:[/*#__PURE__*/e(a,{href:\"https://arxiv.org/abs/2309.06135\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:\"Prompting4Debugging\"})}),' is a specific tool to identify problematic prompts for diffusion models like Stable Diffusion. P4D demonstrated that many prompts, initially deemed \"safe,\" can still bypass existing safeguards.']}),/*#__PURE__*/e(\"img\",{alt:\"\",className:\"framer-image\",height:\"176\",src:\"https://framerusercontent.com/images/q66MJsZI4ebhcROhyCMooF1gvo.png\",srcSet:\"https://framerusercontent.com/images/q66MJsZI4ebhcROhyCMooF1gvo.png?scale-down-to=512 512w,https://framerusercontent.com/images/q66MJsZI4ebhcROhyCMooF1gvo.png 841w\",style:{aspectRatio:\"841 / 352\"},width:\"420\"}),/*#__PURE__*/t(\"p\",{children:[/*#__PURE__*/e(\"em\",{children:\"The Prompting4Debugging framework employs prompt engineering techniques to red-team the text-to-image diffusion model (Source:\"}),/*#__PURE__*/e(a,{href:\"https://arxiv.org/pdf/2309.06135\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:/*#__PURE__*/e(\"em\",{children:\" Prompting4Debugging\"})})}),/*#__PURE__*/e(\"em\",{children:\")\"})]}),/*#__PURE__*/e(\"p\",{children:\"Together, all these tools enable a thorough examination of generative AI models. By integrating relevant resources into red teaming processes, organizations can better safeguard their AI systems, contributing to developing safer and more trustworthy AI technologies.\"}),/*#__PURE__*/e(\"h2\",{children:\"How to make your model safe with Toloka\"}),/*#__PURE__*/t(\"p\",{children:[\"AI red teaming by Toloka can ensure your AI model is robust against potential threats and aligned with industry best practices. Our team of AI researchers, linguists, and subject matter specialists creates prompts specifically designed to trigger unwanted responses from LLMs. We generate a comprehensive report that details the prompts, responses, and classifications to help identify potential safety risks. Additionally, we assist in addressing these issues by delivering high-quality datasets to fine-tune your model, enhancing its safety and ethical standards. \",/*#__PURE__*/e(\"em\",{children:\"Book a\"}),/*#__PURE__*/e(a,{href:\"https://toloka.ai/blog/red-teaming/talk-to-us/\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:/*#__PURE__*/e(\"em\",{children:\" demo\"})})}),/*#__PURE__*/e(\"em\",{children:\" with us to learn more.\"})]}),/*#__PURE__*/t(\"h6\",{children:[\"Curious to learn more? Discover how red teaming enhances both the security and functionality of your AI models in our comprehensive guide\u2014\",/*#__PURE__*/e(a,{href:\"https://toloka.ai/red-teaming-guide?utm_source=blog_article&utm_medium=banner&utm_campaign=red_teaming \",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!0,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:\"click here\"})}),\" to dive in.\"]})]});export const richText2=/*#__PURE__*/t(i.Fragment,{children:[/*#__PURE__*/e(\"p\",{children:\"As large language models (LLMs) transform the AI landscape, AI teams are navigating the\\xa0 GenAI frontier with new approaches to model training and new expectations for fine-tuning data.\\xa0\"}),/*#__PURE__*/e(\"p\",{children:\"Supervised fine-tuning (SFT) is an essential method for training large language models (LLMs) to solve complex problems in niche domains. But to be successful, SFT datasets must meet stringent requirements for quality, expertise, compliance, complexity, and diversity.\\xa0\"}),/*#__PURE__*/e(\"p\",{children:\"Obtaining the appropriate fine-tuning data for your model can sometimes feel like a quest for the holy grail. Let\u2019s look at what a high-quality SFT dataset entails.\"}),/*#__PURE__*/e(a,{href:\"https://toloka.ai/large-language-models/?utm_source=blog&utm_medium=banner&utm_campaign=sft&utm_content=genaifrontier\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!0,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{className:\"framer-image\",\"data-preset-tag\":\"img\",children:/*#__PURE__*/e(\"img\",{alt:\"\",className:\"framer-image\",height:\"240\",src:\"https://framerusercontent.com/images/xpUvAvgGNUA6gpXxlqHiClcP1U.png\",srcSet:\"https://framerusercontent.com/images/xpUvAvgGNUA6gpXxlqHiClcP1U.png?scale-down-to=512 512w,https://framerusercontent.com/images/xpUvAvgGNUA6gpXxlqHiClcP1U.png?scale-down-to=1024 1024w,https://framerusercontent.com/images/xpUvAvgGNUA6gpXxlqHiClcP1U.png 1498w\",style:{aspectRatio:\"1498 / 480\"},width:\"749\"})})}),/*#__PURE__*/e(\"h2\",{children:\"Why is high-quality SFT data so important?\"}),/*#__PURE__*/e(\"p\",{children:\"In the age of generative AI, most models are pre-trained on copious amounts of generic data. Since foundation models are already good at handling basic questions, general human knowledge is not enough to make an LLM stand out. Consider an LLM based on a foundation model \u2014 it can handle a wide range of tasks, but it probably can\u2019t answer a user\u2019s questions about a nuanced topic, like the intricate details of EU law. This is where fine-tuning comes into play.\"}),/*#__PURE__*/e(\"p\",{children:\"Supervised fine-tuning uses a curated dataset focused on a particular skill or area to adapt the LLM to a downstream task and achieve better performance. The model learns from positive examples of desirable answers to prompts, so for a model focused on law, this would entail a large set of complex legal questions and well-written answers. SFT datasets must include specialized, domain-specific knowledge, and the data must be original and unique.\"}),/*#__PURE__*/e(\"h2\",{children:\"Level 1: The quest for expertise\"}),/*#__PURE__*/e(\"p\",{children:\"Expert-level fine-tuning data requires input from knowledgeable subject matter experts. Going back to the example of law, if we\u2019re building a dataset to train an LLM to answer legal questions in the European Union, we want to collect contributions from professionals with a law degree and some experience practicing in the EU. These experts can apply their real-world industry experience to write conversations between an AI agent and a user, sharing their knowledge and showing the model what a truly helpful answer looks like.\\xa0\"}),/*#__PURE__*/t(\"p\",{children:[\"To recruit and vet expert talent, we established the Mindrift platform in 2023. Building on Toloka\u2019s rich 10-year history developing scalable data labeling technologies, the new platform supports a global community of freelance writers, editors, and domain experts \u2014 \",/*#__PURE__*/e(a,{href:\"https://toloka.ai/blog/domain-experts/\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:\"our AI Tutors\"})}),\". Their expertise covers more than 20 domains, including law, healthcare, STEM, compliance, coding, physics, humanities, and more.\\xa0\"]}),/*#__PURE__*/e(\"p\",{children:\"Mindrift attracts a global network of highly educated professionals to work on complex tasks in a collaborative environment. In contrast to anonymous crowdsourcing, AI Tutors support and mentor each other and are dedicated to seeing a project through from start to finish, ensuring consistent results.\\xa0\"}),/*#__PURE__*/e(\"h2\",{children:\"Level 2: The quest for quality\"}),/*#__PURE__*/e(\"p\",{children:\"While quality is paramount for SFT datasets, it depends heavily on the qualification of the experts who generate the data. Our AI Tutors are vetted and tested to rate their skills in writing, fact checking, and generating ethical AI responses. They are assigned to projects in their domain of expertise, with project-specific training to prepare them to do their best work. Automated training and onboarding pipelines help experts get up to speed quickly, so they can start completing tasks with confidence.\\xa0\"}),/*#__PURE__*/e(\"p\",{children:\"Data pipelines include multiple stages that involve the whole team for best quality: each dialog is written, then edited, and verified by experts. A portion of the final data is checked by an auditor to monitor quality. Along the way, experts can discuss difficult tasks with the team and get direct feedback on their submitted tasks to hone their skills. Collaboration within the project team helps reduce fraud and cheating and encourages professional growth.\"}),/*#__PURE__*/e(\"p\",{children:\"To further support our AI Tutors, Mindrift projects use LLMs and other forms of automation to assist with routine steps, so experts can focus on sharing their unique knowledge. For instance, AI-generated prompts offer suggestions to help the expert start writing a dialog. Built-in tools help check grammar, task requirements, factuality, plagiarism, and more.\"}),/*#__PURE__*/e(\"p\",{children:\"Every data production pipeline uses a combination of automated and manual quality checks. But ultimately, it\u2019s the in-depth knowledge and experience of each contributing expert that makes the data valuable for model tuning with SFT.\"}),/*#__PURE__*/e(\"h2\",{children:\"Level 3: The quest for compliance\"}),/*#__PURE__*/e(\"p\",{children:\"Beyond the quality and expertise of the texts, data compliance requirements can be challenging. We need to guarantee that the data submitted by experts is unique and reliable \u2014 it is not plagiarized or under copyright, has not been used in other datasets, and is not AI-generated.\\xa0\"}),/*#__PURE__*/t(\"p\",{children:[\"The Mindrift platform has built-in tools in the task interface to check for plagiarism and detect LLM-generated content, leveraging a combination of detection methods for best accuracy. To learn more about how and why we do this, read our recent article on \",/*#__PURE__*/e(a,{href:\"https://toloka.ai/blog/ai-detection/\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:\"detecting AI-generated texts\"})}),\".\"]}),/*#__PURE__*/e(\"p\",{children:\"In addition, our robust anti-fraud system includes a wide range of proprietary methods to ensure that submitted texts are genuine and original, and our experts are who they say they are.\\xa0\"}),/*#__PURE__*/e(\"h2\",{children:\"Level 4: The quest for complexity and diversity\"}),/*#__PURE__*/e(\"p\",{children:\"Gains in model performance from supervised fine-tuning depend largely on the dataset used \u2014 the size, quality, and diversity of the data, and the length of responses.\\xa0\"}),/*#__PURE__*/e(\"p\",{children:\"Diversity is important for exposing the model to a variety of natural human texts. A balanced dataset includes diverse subdomains, contributors, prompt types, and length and complexity of dialogs.\\xa0\"}),/*#__PURE__*/e(\"p\",{children:\"One way to help maintain diversity is by generating seed topics across a variety of subdomains to guide experts and inspire them. The generated topics also help to vary the prompt type, length, and complexity. In Mindrift, we assign a large pool of experts to each project and track task submissions to encourage diverse contributions and prevent scenarios where a handful of people do all the work.\"}),/*#__PURE__*/e(\"p\",{children:\"No two fine-tuning projects are alike. Some SFT projects may focus more on response length, while others depend on prompt variety or response style. The goals for the model dictate the requirements for the dataset. Non-text modalities \u2014 image, video, and audio data \u2014 are also in demand, along with multi-modal datasets.\\xa0\"}),/*#__PURE__*/e(\"h2\",{children:\"How we build SFT datasets\"}),/*#__PURE__*/e(\"p\",{children:\"While datasets for GenAI are becoming increasingly complex, data production needs are the same as ever: high-quality data at a large scale, delivered in short timeframes and on budget to accelerate AI development.\"}),/*#__PURE__*/t(\"p\",{children:[\"The Toloka team develops \",/*#__PURE__*/e(a,{href:\"https://toloka.ai/blog/crafting-your-own-synthetic-data-pipeline-for-sft/\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:\"custom datasets for fine-tuning (SFT)\"})}),\" that include human-generated prompts and answers. We also offer data collection for alignment (RLHF) and model evaluation.\\xa0\"]}),/*#__PURE__*/e(\"p\",{children:\"To achieve scalable data production, we build highly automated multi-step pipelines that turn implicit human knowledge into structured high-quality datasets. Here are some highlights of our data production pipelines for GenAI:\\xa0\"}),/*#__PURE__*/t(\"ul\",{children:[/*#__PURE__*/e(\"li\",{\"data-preset-tag\":\"p\",style:{\"--framer-font-size\":\"11px\",\"--framer-text-color\":\"rgb(0, 0, 0)\",\"--framer-text-decoration\":\"none\"},children:/*#__PURE__*/e(\"p\",{children:\"Synthetic prompts generated as a starting point to inspire domain experts to write detailed dialogs\"})}),/*#__PURE__*/e(\"li\",{\"data-preset-tag\":\"p\",style:{\"--framer-font-size\":\"11px\",\"--framer-text-color\":\"rgb(0, 0, 0)\",\"--framer-text-decoration\":\"none\"},children:/*#__PURE__*/e(\"p\",{children:\"LLM co-pilots in the task interface help experts work efficiently\"})}),/*#__PURE__*/e(\"li\",{\"data-preset-tag\":\"p\",style:{\"--framer-font-size\":\"11px\",\"--framer-text-color\":\"rgb(0, 0, 0)\",\"--framer-text-decoration\":\"none\"},children:/*#__PURE__*/e(\"p\",{children:\"Extra eyes on the data: multiple experts involved in writing, editing, and verification for each data point\"})}),/*#__PURE__*/e(\"li\",{\"data-preset-tag\":\"p\",style:{\"--framer-font-size\":\"11px\",\"--framer-text-color\":\"rgb(0, 0, 0)\",\"--framer-text-decoration\":\"none\"},children:/*#__PURE__*/e(\"p\",{children:\"61 proprietary anti-fraud methods\"})}),/*#__PURE__*/e(\"li\",{\"data-preset-tag\":\"p\",style:{\"--framer-font-size\":\"11px\",\"--framer-text-color\":\"rgb(0, 0, 0)\",\"--framer-text-decoration\":\"none\"},children:/*#__PURE__*/e(\"p\",{children:\"Over 40 quality control methods\"})})]}),/*#__PURE__*/e(\"p\",{children:\"With the right partnerships, the quest for SFT data doesn\u2019t have to be long and difficult. The Toloka team\u2019s experienced dataset architects partner with each client to define requirements and build efficient data production pipelines. With a solid foundation of human insight and technology, we achieve a final dataset that caters to the model\u2019s exact fine-tuning needs.\"}),/*#__PURE__*/t(\"p\",{children:[/*#__PURE__*/e(a,{href:\"https://toloka.ai/talk-to-us\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!0,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:\"Talk to our dataset architects\"})}),\" about SFT\"]})]});export const richText3=/*#__PURE__*/t(i.Fragment,{children:[/*#__PURE__*/e(\"p\",{children:\"Large language models (LLMs) can understand and generate human-like text, making them incredibly useful for various applications, from chatbots to content creation. Their deployment on the cloud significantly enhances their accessibility and efficiency, making it easier to leverage their capabilities.\"}),/*#__PURE__*/e(\"p\",{children:\"A cloud-based large language model works in a cloud environment, so investing in expensive hardware and infrastructure to host and operate is unnecessary. Businesses can utilize the powerful computing resources provided by the cloud providers without purchasing and maintaining their own data centers. In this article, we\u2019ll explore cloud-based LLMs, how they work, and why they matter.\"}),/*#__PURE__*/e(\"h2\",{children:\"What are cloud LLMs?\"}),/*#__PURE__*/e(\"p\",{children:\"A large language model is an artificial intelligence (AI) model designed to understand and generate text. These models are typically built using deep learning techniques and trained on massive data. Training and running LLMs require substantial computational resources, which can be costly and environmentally impactful. Due to budget constraints, many enterprises need help to afford to deploy and fine-tune LLM models.\"}),/*#__PURE__*/e(\"p\",{children:\"Cloud computing involves delivering computing services\u2014servers, storage, databases, networking, software, and more\u2014online (in the cloud). It has revolutionized how we use and interact with technology, and LLMs are no exception. Major cloud service providers like Amazon Web Services (AWS), Microsoft Azure, and Google Cloud Platform offer the infrastructure and tools necessary to run large-scale applications without needing physical hardware on-premises.\"}),/*#__PURE__*/e(\"p\",{children:\"By deploying LLMs on cloud platforms, companies can leverage the immense computational power needed to train and run these models without investing in expensive hardware. Users can access these models via APIs, making it easy to integrate advanced language processing capabilities into their applications with minimal effort.\"}),/*#__PURE__*/e(\"h3\",{children:\"LLM as a service\"}),/*#__PURE__*/t(\"p\",{children:[\"Large companies tend to run most of their applications locally, meaning the systems and databases they use reside on their own servers. According to the \",/*#__PURE__*/e(a,{href:\"https://www.mckinsey.com/capabilities/mckinsey-digital/our-insights/in-search-of-cloud-value-can-generative-ai-transform-cloud-roi\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:\"McKinsey report\"})}),\", cloud services most often take up only 15% to 20% of the IT infrastructure, even though reliable cloud services have been around for a long time and offer a wide range of business benefits. The computational power and expertise required to deploy and manage large language models can be a significant barrier for enterprises. In other words, LLM as a service could be the factor that accelerates cloud adoption by businesses. It refers to cloud-based platforms that offer access to LLMs via APIs.\\xa0\"]}),/*#__PURE__*/e(\"p\",{children:\"These services allow developers and businesses to integrate the capabilities of sophisticated language models into their applications without managing the whole AI infrastructure. Moreover, companies can pay only for what they use, avoiding the upfront costs of purchasing and maintaining high-performance hardware.\"}),/*#__PURE__*/e(\"p\",{children:\"Users do not need to worry about the complexities of maintaining the language models. Cloud-based LLMs enable non-technical specialists to benefit from advanced machine learning capabilities without deep technical knowledge.\"}),/*#__PURE__*/e(\"p\",{children:\"These platforms offer user-friendly interfaces, APIs, and pre-trained models that simplify the integration of sophisticated natural language processing (NLP) capabilities into various applications. For instance, businesses can now automate customer support, generate content, analyze sentiment, and more with minimal technical intervention.\"}),/*#__PURE__*/e(\"h2\",{children:\"What are local LLMs?\"}),/*#__PURE__*/e(\"p\",{children:\"Local large language models refer to large-scale language models that are deployed and run on local hardware, such as on-premises data centers, that do not rely on cloud-based infrastructure. They are usually more suitable for companies with AI research labs and teams with high ML expertise.\"}),/*#__PURE__*/e(\"p\",{children:\"When deploying large language models locally, specialists have more control over the model, which makes more precise customization possible. This can include fine-tuning the model on specific datasets to improve performance for particular tasks or industries. Although the set-up costs may be higher initially due to the need for powerful hardware and the costs of expertise, on-premises LLMs are potentially more cost-effective in the long run due to the absence of ongoing fees for cloud platforms.\"}),/*#__PURE__*/e(\"p\",{children:\"However, scaling up local LLMs to handle larger volumes of data or more complex tasks can be challenging compared to scalable cloud solutions. More than that, technical expertise and regular maintenance are required to keep the LLM models updated and make them run efficiently.\"}),/*#__PURE__*/e(\"h2\",{children:\"Pros of using LLMs in the cloud\"}),/*#__PURE__*/t(\"p\",{children:[/*#__PURE__*/e(\"strong\",{children:\"Cost Efficiency. \"}),\"The pay-as-you-go pricing model of cloud services allows for cost management based on actual usage. This reduces the need for substantial upfront investment in hardware and can be more economical, especially for projects with variable computational needs;\"]}),/*#__PURE__*/t(\"p\",{children:[/*#__PURE__*/e(\"strong\",{children:\"Scalability.\"}),\" Cloud-based LLMs can handle varying workloads, from small projects to enterprise-level applications, by efficiently scaling resources up or down based on demand. This flexibility ensures that you can manage peak loads without investing in additional hardware;\"]}),/*#__PURE__*/t(\"p\",{children:[/*#__PURE__*/e(\"strong\",{children:\"Accessibility.\"}),\" Cloud LLMs are accessible from anywhere with an internet connection. This makes it easier for distributed teams to collaborate and for applications to be deployed globally without geographical constraints;\"]}),/*#__PURE__*/t(\"p\",{children:[/*#__PURE__*/e(\"strong\",{children:\"Maintenance-Free.\"}),\" The cloud service provider maintains the infrastructure, including updates, patches, and hardware maintenance. This reduces the burden on internal IT resources and ensures that the models always run on the latest and most secure platforms.\"]}),/*#__PURE__*/e(\"h2\",{children:\"Cons of using LLMs in the cloud\"}),/*#__PURE__*/t(\"p\",{children:[/*#__PURE__*/e(\"strong\",{children:\"Data Privacy and Security Concerns.\"}),\" Sensitive data needs to be transmitted to and from the cloud, which can pose risks of data breaches and raise concerns about compliance with data protection regulations;\"]}),/*#__PURE__*/t(\"p\",{children:[/*#__PURE__*/e(\"strong\",{children:\"Speed Issues. \"}),\"Depending on the network speed and location of the cloud servers, there can be latency in processing requests. This can be a challenge for applications that need real-time responses;\"]}),/*#__PURE__*/t(\"p\",{children:[/*#__PURE__*/e(\"strong\",{children:\"Ongoing Costs.\"}),\" While the pay-as-you-go model can be cost-efficient, it can also lead to significant ongoing costs, especially for high-volume usage. Over time, these costs may accumulate and potentially exceed the cost of running LLMs locally;\"]}),/*#__PURE__*/t(\"p\",{children:[/*#__PURE__*/e(\"strong\",{children:\"Dependency on Internet Connectivity.\"}),\" Cloud-based solutions require a stable internet connection. Any disruption in connectivity can affect access to the LLMs and disrupt services;\"]}),/*#__PURE__*/t(\"p\",{children:[/*#__PURE__*/e(\"strong\",{children:\"Limited Customization. \"}),\"While cloud providers offer some customization, there may be limitations compared to what can be achieved with local deployments. This can be a drawback for specialized applications requiring deep integration or specific optimizations;\"]}),/*#__PURE__*/e(\"h2\",{children:\"Local vs. Cloud LLMs\"}),/*#__PURE__*/e(\"p\",{children:\"Local LLMs offer significant advantages in terms of data privacy and security,\\xa0 running models locally ensures that sensitive data remains within the internal infrastructure. Unlike cloud services, deploying the LLMs locally reduces the risk of data breaches and helps to comply with strict data protection regulations. In addition, local models reduce the latency associated with sending data to and receiving data from remote servers, thus making them ideal for applications that require quick response.\"}),/*#__PURE__*/e(\"p\",{children:\"However, local LLMs have several drawbacks. The initial setup costs can be significant, requiring investment in hardware and infrastructure, such as powerful GPUs and storage solutions. They also necessitate ongoing maintenance, updates, and potentially additional staffing or expertise to manage the infrastructure and software. Moreover, scaling up may involve additional hardware purchases and setup, which can be less flexible compared to cloud solutions.\"}),/*#__PURE__*/e(\"p\",{children:\"On the other hand, cloud LLMs offer their own set of advantages. They are easily scalable to handle varying workloads, from small-scale to enterprise-level applications, and can quickly adjust resources as needed. This scalability is coupled with lower initial costs since there\u2019s no need to purchase expensive hardware, and pay-as-you-go pricing models allow for cost management based on usage.\"}),/*#__PURE__*/e(\"p\",{children:\"Cloud LLMs are also maintenance-free, as the service provider handles model updates, infrastructure maintenance, and performance tuning, freeing internal resources. Additionally, cloud-based models are accessible from anywhere with an internet connection, facilitating remote work and collaboration.\"}),/*#__PURE__*/e(\"p\",{children:\"Despite these benefits, cloud LLMs have their disadvantages. There are data privacy and security concerns, as sensitive data must be transmitted to and from the cloud. This can raise concerns about data breaches and compliance with data protection laws. Potential latency due to network delays can also be problematic for real-time applications. Lastly, ongoing operational costs, based on usage, can accumulate over time and potentially exceed the cost of local infrastructure.\"}),/*#__PURE__*/e(\"h2\",{children:\"Cloud or local LLM hosting: what to choose?\"}),/*#__PURE__*/e(\"p\",{children:\"The emergence of cloud-based large language models is a groundbreaking development that significantly lowers the barriers to utilizing advanced machine learning models. Cloud platforms make LLMs accessible to many enterprises, including those with limited technical expertise. A cloud provider takes care of updates, support, and security, reducing the burden on your team. However, cloud solutions have drawbacks. Transmitting sensitive data over the internet can be a security risk, and delays can occur due to network and data transfer speed.\"}),/*#__PURE__*/e(\"p\",{children:\"Locally hosted LLMs also have their advantages. The most important of these is that data is processed locally, which reduces the risk of data breaches. The models perform faster since there is no need to transfer data online, the models perform at faster speeds. Being locally hosted provides more opportunities to optimize models for specific tasks. But this approach also has its drawbacks. It requires significant amount of computing resources, such as powerful GPUs and RAM. Hardware and software require regular upgrades and maintenance. If additional capacity is needed, it will demand extra investment in hardware.\"}),/*#__PURE__*/e(\"p\",{children:\"Consequently, the choice between cloud and on-premises hosting of large language models depends on factors such as your business's needs and available resources. Cloud hosting is suitable for companies that are looking for scalability and no major upfront costs, while on-premises hosting is better for organizations with high-security requirements and those willing to invest in powerful hardware and technical support. Evaluate your priorities and resources to make the most appropriate choice for your company.\"})]});export const richText4=/*#__PURE__*/t(i.Fragment,{children:[/*#__PURE__*/e(\"p\",{children:\"Language models are essential for enabling machines to understand and generate human language. Large Language Models (LLMs) often receive the most attention, boasting billions of parameters and excelling across various tasks. However, a new trend is emerging with small language models (SLMs). These smaller models strike a balance between computational power and efficiency, making artificial intelligence (AI) more accessible and widely adopted.\"}),/*#__PURE__*/e(\"h2\",{children:\"What are Language Models?\"}),/*#__PURE__*/e(\"p\",{children:\"A language model is an algorithm that calculates the probability for each word in a language to occur in a particular context. Because there are so many words in any language, the model is taught to compute probabilities only for words in a particular vocabulary,which is a relatively small set of words or parts of words in a language.\"}),/*#__PURE__*/e(\"p\",{children:\"In simpler terms, language models can continue texts. The model calculates the probability of possible continuations of a text and suggests them. It assigns probabilities to sequences of words and predicts the next word in a sentence given the previous words. Its main goal is to understand the structure and patterns of language to generate coherent and contextually appropriate text.\"}),/*#__PURE__*/e(\"p\",{children:\"Transformers are a fundamental architecture in modern natural language processing that has radically reshaped how models work with sequential data. The main innovation of transformers is the self-attention mechanism, which allows the model to evaluate the importance of different words in a sentence relative to each other.\"}),/*#__PURE__*/e(\"p\",{children:\"Why is the development of modern AI models centered around language models? Indeed, language models are the core of modern natural language processing (NLP) and artificial intelligence. Their existence and such rapid development are driven by the need to enable machines to understand, interpret, and generate human language.\"}),/*#__PURE__*/e(\"p\",{children:\"Researchers started developing language models because human language is a fundamental to intelligence and communication. Practical needs for improved communication, information search, and automation, combined with the historical context of AI research and technological advances in machine learning and data availability, have prioritized language models.\"}),/*#__PURE__*/e(\"h2\",{children:\"Types of Language Models\"}),/*#__PURE__*/e(\"h3\",{children:\"Large Language Models\"}),/*#__PURE__*/e(\"p\",{children:\"Large language models (LLMs), such as GPT-3 with 175 billion parameters or BERT with 340 million parameters, are designed to perform highly in all kinds of natural language processing tasks. Parameters are variables of a model that change during the learning process. The large language model is a neural linguistic network trained on extensive and diverse datasets, which allows it to understand complex language patterns and long-range dependencies.\"}),/*#__PURE__*/e(\"p\",{children:\"Such highly versatile models can be fine-tuned to become domain-specific language models. LLMs are great for various complex tasks, from text generation and translation to summarization and advanced research tasks. However, LLMs require significant computational resources, memory, and storage, making them expensive to train and deploy. They also consume a lot of energy and have slower inference times, which can be a drawback for real-time applications.\"}),/*#__PURE__*/e(\"h3\",{children:\"Small Language Models\"}),/*#__PURE__*/e(\"p\",{children:\"Small language models are neural networks designed to perform natural language processing tasks with a significantly reduced number of parameters and with fewer computational resources than larger language models.\"}),/*#__PURE__*/e(\"p\",{children:\"Small models are trained on more limited datasets and often use techniques like knowledge distillation to retain the essential features of larger models while significantly reducing their size. This makes them much more cost-effective to train and deploy even on mobile devices because they require less computational power and storage. Their faster inference times make them suitable for real-time applications like chatbots and mobile apps.\"}),/*#__PURE__*/e(\"p\",{children:\"Small language models, such as DistilBERT with 66 million parameters or TinyBERT with approximately 15 million parameters, are optimized for efficiency. They are trained on more specific datasets or subsets of larger corpora. Being trained on limited datasets, small models often use techniques like distillation to retain the essential features of larger models while significantly reducing their size. Capable small language models are more accessible than their larger counterparts to organizations with limited resources, including smaller organizations and individual developers.\"}),/*#__PURE__*/e(\"p\",{children:\"While LLMs might have hundreds of billions of parameters, SLMs typically operate with parameters in the millions to low billions range. In contrast to all-purpose LLMs, small models are designed for highly specialized tasks as they handle them noticeably better. Despite their smaller size, these models ca be remarkably effective, especially for specific tasks or when optimized using advanced training techniques.\"}),/*#__PURE__*/e(\"p\",{children:\"There has yet to be\\xa0 clearly defined distinction between LLM and SLM yet. Likewise, there is no clear definition of hthe number ofparameters large and small language models have. Larger models are considered to handle 100 million or more parameters, or according to other sources, 100+ billion. Small language models are considered to handle fewer parameters ranging from 1 to 10 million, or 10 billion.\"}),/*#__PURE__*/e(\"p\",{children:\"Large and small language models differ not only in the number of parameters, but also in the amount of data processed, training data, required storage, and neural architecture. A small language model requires significantly fewer resources to train and analyze data than a large one.\"}),/*#__PURE__*/e(\"h4\",{children:\"Comparative Summary: Small Language Models Vs. Large Language Models\"}),/*#__PURE__*/e(\"img\",{alt:\"\",className:\"framer-image\",height:\"392\",src:\"https://framerusercontent.com/images/OZzBEnwycY3UDhaKrsTQJm6YY.jpeg\",srcSet:\"https://framerusercontent.com/images/OZzBEnwycY3UDhaKrsTQJm6YY.jpeg?scale-down-to=512 512w,https://framerusercontent.com/images/OZzBEnwycY3UDhaKrsTQJm6YY.jpeg?scale-down-to=1024 1024w,https://framerusercontent.com/images/OZzBEnwycY3UDhaKrsTQJm6YY.jpeg 1394w\",style:{aspectRatio:\"1394 / 784\"},width:\"697\"}),/*#__PURE__*/e(\"h2\",{children:\"Why Are SLMs Necessary?\"}),/*#__PURE__*/e(\"p\",{children:\"Small language models are significant for several reasons. Firstly, many devices we use daily - smartphones, tablets, and even items like smart home gadgets - don't possess much processing power. Small language models only need a little processing power, memory, or storage, so they work great in these environments. They keep devices running smoothly without draining their resources.\"}),/*#__PURE__*/e(\"p\",{children:\"Cost is another critical factor. Large language models are costly to train and use because they require a lot of computing power. Small models are much cheaper to run, meaning that cutting-edge NLP becomes affordable for more companies and developers, even with limited budgets. They also consume fewer resources, lowering operating costs and reducing environmental impact.\"}),/*#__PURE__*/e(\"h2\",{children:\"Benefits of Small Language Models\"}),/*#__PURE__*/t(\"p\",{children:[/*#__PURE__*/e(\"strong\",{children:\"Lower computational requirements\"}),\". Small language models require significantly less computational power and memory compared to large language models. This makes them more accessible for use on devices with limited resources, like smartphones, tablets, and edge devices.\"]}),/*#__PURE__*/t(\"p\",{children:[/*#__PURE__*/e(\"strong\",{children:\"Reduced training and inference costs\"}),\". Due to their smaller size, these models are cheaper to train and deploy. This cost-effectiveness benefits organizations with limited budgets or those looking to deploy NLP solutions at scale without high infrastructure costs.\"]}),/*#__PURE__*/t(\"p\",{children:[/*#__PURE__*/e(\"strong\",{children:\"Speed. \"}),\"With fewer parameters, small language models have faster inference times. This is crucial for real-time applications such as chatbots, virtual assistants, and mobile applications, where quick responses are essential for a good user experience. Small models can also be fine-tuned more quickly with less data compared to larger models. This adaptability is helpful for domain-specific applications where training data might be limited, and rapid iteration is necessary.\"]}),/*#__PURE__*/t(\"p\",{children:[/*#__PURE__*/e(\"strong\",{children:\"Deployment in resource-constrained environments\"}),\". Small language models can be easily deployed in environments with constrained computational resources. This includes IoT devices, embedded systems, and other edge cases where large models would be impractical. Small language models' reduced size and complexity of small language models make them easier to deploy on various platforms, including mobile devices and embedded systems. This flexibility benefits applications that need to run in environments with different hardware capabilities.\"]}),/*#__PURE__*/t(\"p\",{children:[/*#__PURE__*/e(\"strong\",{children:\"Wider accessibility.\"}),\" Lower costs and reduced hardware requirements make small language models more accessible to small organizations, academic institutions, and even individual developers. This contributes to broader access to advanced NLP technologies, allowing a wider range of stakeholders to benefit from AI breakthroughs.\"]}),/*#__PURE__*/e(\"h2\",{children:\"Small Language Models Use Cases\"}),/*#__PURE__*/e(\"h3\",{children:\"Customer Support and Chatbots\"}),/*#__PURE__*/e(\"p\",{children:\"SLMs are used to develop AI-powered chatbots and virtual assistants that handle customer inquiries, provide personalized responses, and automate routine tasks. Their efficient understanding and generating natural language makes them ideal for enhancing customer service experiences.\"}),/*#__PURE__*/e(\"h3\",{children:\"Sentiment Analysis and Social Media Monitoring\"}),/*#__PURE__*/e(\"p\",{children:\"SLMs analyze text data from social media platforms and online forums to perform sentiment analysis, identifying trends, public opinion, and potential issues. This capability helps businesses and organizations monitor their online presence and customer sentiment.\"}),/*#__PURE__*/e(\"h3\",{children:\"Language Translation and Localization\"}),/*#__PURE__*/e(\"p\",{children:\"SLMs contribute to language translation services by accurately translating text between languages, improving accessibility to information across global audiences. They can handle nuances in language and context, facilitating effective communication in multilingual environments.\"}),/*#__PURE__*/e(\"h3\",{children:\"Content Generation and Summarization\"}),/*#__PURE__*/e(\"p\",{children:\"In media and publishing, SLMs are employed for content-generation tasks such as writing articles, generating product descriptions, and creating summaries of long documents or reports. They can produce coherent and contextually relevant content quickly and efficiently.\"}),/*#__PURE__*/e(\"h2\",{children:\"Best small language models\"}),/*#__PURE__*/e(\"p\",{children:/*#__PURE__*/e(\"strong\",{children:\"LLaMA 3\"})}),/*#__PURE__*/e(\"p\",{children:\"LLaMA 3 is open-source, which means its code and architecture are accessible to the public. Developed by Meta, LLaMA 3 with 8 billion parameters is part of their strategy to empower the AI community with trustworthy tools. Meta aims to foster responsible AI usage by providing models that are not only powerful but also adaptable to various applications and scenarios. By making LLaMA 3 open-source, Meta encourages collaboration and contributions from researchers, developers, and enthusiasts worldwide.Phi-3\"}),/*#__PURE__*/e(\"p\",{children:\"Phi-3 represents Microsoft's commitment to advancing AI accessibility by offering powerful yet cost-effective solutions. Phi-3 models are part of Microsoft\u2019s open AI initiative, emphasizing transparency and accessibility. This means that the models are publicly accessible, allowing developers and researchers to integrate them into different environments. Despite its smaller size compared to many other models in the field, Phi-3, with 3.8 billion parameters,has demonstrated superior performance in benchmark tests.\"}),/*#__PURE__*/e(\"p\",{children:/*#__PURE__*/e(\"strong\",{children:\"Gemma\"})}),/*#__PURE__*/e(\"p\",{children:\"Google's Gemma family of AI models represents a significant advancement in open-source AI technology. Gemma models are part of Google's initiative to provide open-source AI solutions that are accessible to developers worldwide. This openness allows developers to explore, modify, and integrate the models into their applications with greater freedom and control. Gemma models are engineered to be lightweight, making them suitable for deployment on various devices and platforms.\"}),/*#__PURE__*/e(\"h2\",{children:\"Small Language Models Challenges\"}),/*#__PURE__*/e(\"p\",{children:\"SLMs are optimized for specific tasks or domains, which often allows them to operate more efficiently regarding computational resources and memory usage compared to larger models. By focusing on a narrow domain, efficient small language models can achieve higher accuracy and relevance within their specialized area.\"}),/*#__PURE__*/e(\"p\",{children:\"However, SLMs may lack the broad knowledge base necessary to generalize well across diverse topics or tasks. This limitation can reduce performance or relevance when applied outside their trained domain. Moreover, organizations may need to deploy multiple SLMs, each specialized in different domains or tasks, to effectively cover a wide range of needs effectively. Managing and integrating these models into a cohesive AI infrastructure can be resource-intensive.\"}),/*#__PURE__*/e(\"p\",{children:\"Although niche-focused SLMs offer efficiency advantages, their limited generalization capabilities require careful consideration. A balance between these compromises is necessary to optimize the AI infrastructure and effectively use both small and large language models.\"}),/*#__PURE__*/e(\"h2\",{children:\"Future of Small Language Models\"}),/*#__PURE__*/e(\"p\",{children:\"SLMs contribute to democratizing AI by making advanced technology more accessible to a broader audience. Their smaller size and efficient design lower barriers to entry for developers, researchers, startups, and communities that may have limited resources or expertise in deploying AI solutions.\"}),/*#__PURE__*/e(\"p\",{children:\"Collaboration among researchers, stakeholders, and communities will drive further innovation in SLMs. Open dialogue, shared resources, and collective efforts are essential to maximizing AI's positive impact on society.\"}),/*#__PURE__*/e(\"p\",{children:\"With the differences between SLM and LLM gradually diminishing, there will appear new ways to apply AI will appear\\xa0 real-world applications. As research progresses, SLMs are expected to become more efficient regarding computational requirements while maintaining or even improving their performance.\"}),/*#__PURE__*/e(\"p\",{children:\"Small language models represent a pivotal advancement in democratizing AI, making it more accessible, adaptable, and beneficial to many users and applications. As technology evolves and barriers diminish, SLMs will continue to shape a future with AI enhancing human capabilities effectively.\"})]});export const richText5=/*#__PURE__*/t(i.Fragment,{children:[/*#__PURE__*/t(\"p\",{children:[\"You've most likely heard about GPT-4 and its remarkable capabilities. In this article, we run through 155 pages of the research paper \",/*#__PURE__*/e(a,{href:\"https://arxiv.org/abs/2303.12712\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:'\"Sparks of AGI'})}),',\" in which Microsoft researchers present one of OpenAI\\'s latest technology advancements, an early version of GPT-4. It convincingly demonstrates remarkable performance in many aspects, such as answering Fermi questions. These questions require logic and approximation to estimate quantities that are hard to measure directly, for example: \"How long would it take to count to one billion?\".\\xa0 Furthermore, it can handle tasks from the physical world, like assisting a human in repairing a water leak. The central claim around this model is that it can be considered early progress towards an artificial general intelligence system. The early experiments showcase GPT-4\\'s ability to solve novel and difficult tasks in various fields, such as coding, mathematics, drawing, music, and even tricky social situations.\\xa0']}),/*#__PURE__*/e(\"p\",{children:\"We review these claims by analyzing the early experiments and discussing the achievements and known limitations.\"}),/*#__PURE__*/e(\"h2\",{children:\"Unveiling GPT-4: How do you measure genius?\"}),/*#__PURE__*/e(\"p\",{children:\"A common challenge computer scientists face is assessing the performance of a model trained on an unprecedented amount of text data. Most models are evaluated on independent benchmark datasets across various domains. While this method measures actual learning rather than merely memorization, the authors argue that there are limits to these assessments. There are two main reasons for not choosing such a standard evaluation technique. Since there is limited access to its complete training data, we can assume that GPT-4 has likely encountered most existing benchmarks. More importantly, GPT-4's intelligence is characterized by its generality, allowing it to perform tasks that are out of reach for domain-specific AI systems. Evaluating GPT-4 on generative or interactive tasks is challenging, as these are not single-solution tasks and are difficult to assess.\\xa0\"}),/*#__PURE__*/e(\"p\",{children:\"In light of this, the authors have opted for an approach that aligns more with traditional psychology than machine learning. They aim to leverage human ingenuity and curiosity to demonstrate GPT-4's deep and flexible understanding by testing it on novel and challenging tasks.\"}),/*#__PURE__*/e(\"h2\",{children:\"Milestones: A journey through GPT-4's achievements\\xa0\"}),/*#__PURE__*/e(\"p\",{children:\"Throughout this research paper, the authors present some impressive things GPT-4 can do. They conducted experiments in several domains specifically chosen to analyze its fundamental abilities: learning from experience, capacity to plan, quick learning, and problem-solving.\"}),/*#__PURE__*/e(\"h3\",{children:\"The eyes of artificial intelligence: Visual feats\"}),/*#__PURE__*/e(\"p\",{children:\"Even though GPT-4 is a non-multimodal language model, it can still comprehend simple visual concepts. For example, it can generate Scalable Vector Graphics (SVG) images of four objects: a car, a truck, a cat, and a dog.\\xa0\"}),/*#__PURE__*/e(\"img\",{alt:\"\",className:\"framer-image\",height:\"113\",src:\"https://framerusercontent.com/images/DJQhytsXurAlgcNpOGXwIjneW88.png\",srcSet:\"https://framerusercontent.com/images/DJQhytsXurAlgcNpOGXwIjneW88.png?scale-down-to=512 512w,https://framerusercontent.com/images/DJQhytsXurAlgcNpOGXwIjneW88.png 819w\",style:{aspectRatio:\"819 / 227\"},width:\"409\"}),/*#__PURE__*/t(\"p\",{children:[\"ref:\",/*#__PURE__*/e(a,{href:\"https://arxiv.org/abs/2303.12712\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:\" sparks of artificial general intelligence\"})}),\" (page 16, figure 2.4)\",/*#__PURE__*/e(\"br\",{}),/*#__PURE__*/e(\"br\",{className:\"trailing-break\"})]}),/*#__PURE__*/e(\"img\",{alt:\"\",className:\"framer-image\",height:\"260\",src:\"https://framerusercontent.com/images/In85YMwyeB4RBL1poodko57Qw.png\",srcSet:\"https://framerusercontent.com/images/In85YMwyeB4RBL1poodko57Qw.png?scale-down-to=512 512w,https://framerusercontent.com/images/In85YMwyeB4RBL1poodko57Qw.png 933w\",style:{aspectRatio:\"933 / 520\"},width:\"466\"}),/*#__PURE__*/t(\"p\",{children:[\"ref:\",/*#__PURE__*/e(a,{href:\"https://arxiv.org/abs/2303.12712\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:\" sparks of artificial general intelligence\"})}),\" (page 16, figure 2.5)\"]}),/*#__PURE__*/e(\"p\",{children:\"Even so, similar images could be present in the training data, and one may argue that the model just reproduced the code from there. However, in the following instance, GPT-4 is instructed to combine the letters H, O, and Y to draw a human and even add clothing. It can associate different geometric shapes to a human torso, showcasing a broader grasp of visual notions beyond simply copying existing code.\\xa0\"}),/*#__PURE__*/e(\"h3\",{children:\"Note-worthy: Musical creations\"}),/*#__PURE__*/e(\"p\",{children:\"GPT-4 can even write sheet music in ABC notation, although the outcome is relatively basic and limited. It produces accurate melodies and repetitive rhythm and can explain the overall structure of the tune. Its limitations, however, become apparent in the lack of harmony, where the model demonstrated little to no conceptual understanding. Possible reasoning for this outcome, as stated in the paper, is the lack of widespread adoption of the ABC notation, which would explain why GPT-4 couldn't even recognize well-known melodies, such as Beethoven's Ode to Joy.\"}),/*#__PURE__*/e(\"h3\",{children:\"Bit by bit: Cracking the code\"}),/*#__PURE__*/e(\"p\",{children:\"This section highlights GPT-4's coding capabilities through coding challenges and real-world applications. It demonstrates its proficiency in coding complex tasks, from low-level components to high-level architectures. Additionally, the model can interpret and execute pseudo code, which involves understanding informal and often imprecise expressions unsupported by programming languages.\"}),/*#__PURE__*/e(\"p\",{children:\"When tested against a few benchmarks, GPT-4 significantly outperforms other large language models. It achieves nearly 20% higher accuracy on the HumanEval and LeetCode benchmarks than the second-best model, text-davinci-003 (the base model of ChatGPT). On top of that, GPT-4's performance on the LeetCode benchmark nearly matches human performance, which is superior by only 0.2%.\\xa0\"}),/*#__PURE__*/e(\"p\",{children:\"More examples include developing simple games using HTML and JavaScript, writing incredibly complex functions in LaTeX, and even applying it to deep learning tasks, such as writing a custom PyTorch machine learning optimizer.\"}),/*#__PURE__*/e(\"p\",{children:\"Beyond just writing code, GPT-4 tries to guess a password by reverse-engineering a binary executable code written in C. It uses tools such as GDB for debugging and Python for writing the 'crack-the-password' code. Interestingly, ChatGPT refuses to comply with the exact instructions, claiming that doing so would be unethical, even though reverse engineering is often used to improve software security. GPT-4, on the other hand, compares the password to a hash value derived from a mathematical expression, eventually figuring it out by guessing the right combination of digits that matches the value.\"}),/*#__PURE__*/e(\"h3\",{children:\"Mathematical marvels: How it all adds up\"}),/*#__PURE__*/e(\"p\",{children:\"GPT-4 can solve high-school-level math problems and occasionally explain advanced math topics reasonably. In the same way, it can answer Fermi questions and tackle graph theory and algorithms.\"}),/*#__PURE__*/e(\"p\",{children:\"However, it frequently generates basic mistakes and occasionally produces inconsistent results, which may be seen as a lack of conceptual sense and overall general intelligence. The model often makes arithmetic mistakes that would be a no-brainer for humans to solve, and its performance on the MATH dataset confirms just that. Even 68% of the generated solutions for the arithmetic tasks are incorrect.\\xa0\"}),/*#__PURE__*/e(\"p\",{children:\"Critical reasoning is where GPT-4 sees its most significant shortcomings. As stated in the paper, this is likely a challenge all large language models face, since these models are explicitly trained to predict the next word and lack an inner monolog that looks back to correct their previous mistakes.\\xa0\"}),/*#__PURE__*/e(\"p\",{children:'This limitation can be potentially eliminated by adding more mathematical data to the training set, which also includes the \"thinking process\" of solving a mathematical question and not just the linear relationship between the problem and its solution.'}),/*#__PURE__*/e(\"h3\",{children:\"Wielding the toolkit: Crafting solutions with AI\"}),/*#__PURE__*/e(\"p\",{children:\"The enabled interaction of GPT-4 with external tools has emerged as one of the most notable trends for real-world applications. These resources are a great asset and can fill the gaps where GPT-4 lacks specific capabilities, such as up-to-date world knowledge, arithmetic operations, etc. For instance, it cannot provide a correct answer to questions concerning up-to-date events, such as the current president of the US. Additionally, it cannot solve a simple math equation nor identify the 13th letter in the word supralapsarian. However, GPT-4 has no problem executing these tasks when it can access resources like search engines and APIs.\"}),/*#__PURE__*/e(\"p\",{children:\"Having said that, giving the model access to resources alone is insufficient to fix all of the challenges it may encounter. GPT-4 still needs explicit instructions that indicate whether using external tools is permitted or expected. Furthermore, it cannot always rationally decide when and how to apply tools. For instance, in one session, it uses a web search to find the capital of France, even though it should know it on its own.\"}),/*#__PURE__*/e(\"h3\",{children:\"Understanding the human mind: GPT-4's perspective\"}),/*#__PURE__*/e(\"p\",{children:\"The model also exhibits a high level of theory of mind, which is the ability to recognize and process the mental and emotional states of others and oneself. It is able to interpret a situation from someone else's perspective and give an educated guess about their emotional state.\\xa0\"}),/*#__PURE__*/e(\"p\",{children:\"In the example below, it answers questions about a conversation between two people, explaining why one of them is sad.\\xa0\"}),/*#__PURE__*/e(\"img\",{alt:\"\",className:\"framer-image\",height:\"301\",src:\"https://framerusercontent.com/images/D4jFqI8UdmFU1udWBUO6Olczzs.png\",srcSet:\"https://framerusercontent.com/images/D4jFqI8UdmFU1udWBUO6Olczzs.png?scale-down-to=512 512w,https://framerusercontent.com/images/D4jFqI8UdmFU1udWBUO6Olczzs.png?scale-down-to=1024 1024w,https://framerusercontent.com/images/D4jFqI8UdmFU1udWBUO6Olczzs.png 1566w\",style:{aspectRatio:\"1566 / 602\"},width:\"783\"}),/*#__PURE__*/t(\"p\",{children:[\"ref:\",/*#__PURE__*/e(a,{href:\"https://arxiv.org/abs/2303.12712\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:\" sparks of artificial general intelligence\"})}),\" (page 55, figure 6.2)\"]}),/*#__PURE__*/e(\"h2\",{children:\"Beyond the breaking point: LLM limitations\"}),/*#__PURE__*/e(\"p\",{children:\"In this article, we discussed GPT-4's strengths and weaknesses in the context of different challenges and domains.\"}),/*#__PURE__*/e(\"p\",{children:\"One of the main limitations the authors acknowledge is GPT-4's lack of ability to plan ahead, which they attribute to the autoregressive nature of the LLM. The model's inability to work step by step when solving a problem makes it more challenging to provide a correct answer. Interestingly, though, GPT-4 can plan things; we only have to prompt it. For example, when asked how many prime numbers are between 150 and 250, the zero-shot answer is 13, which is wrong. However, if you ask it to list all the numbers and then return the list size, it outputs the correct solution (18), as it is much easier to count the list items. Moreover, it has issues with text generation, as it seems to have difficulty planning ahead on a longer text (global scale), which is also inherent to its next-word prediction architecture.\\xa0\"}),/*#__PURE__*/e(\"p\",{children:\"Other examples include the unavoidable tendency to hallucinate, generate incorrect information, and make basic arithmetic mistakes.\\xa0\"}),/*#__PURE__*/e(\"p\",{children:\"To estimate roughly, GPT-4 excels when so-called fast-thinking is required, which is automatic and intuitive but entirely exposed to biases and errors. On the other hand, it cannot do slow-thinking, which is organizing the thought process and giving a rational, well-thought-out answer.\"}),/*#__PURE__*/e(\"p\",{children:\"Improvement is also needed in other areas to achieve more general intelligence. These include improving long-term working memory, planning, better conceptualization, and learning from experience.\"}),/*#__PURE__*/e(\"h2\",{children:\"Closing thoughts: Does GPT-4 indeed show sparks of artificial general intelligence?\"}),/*#__PURE__*/e(\"p\",{children:\"This paper's early exploration of GPT-4's capabilities suggests that it performs at a human level for many tasks and domains. One may wonder if GPT-4 truly grasps the explored concepts or simply excels at improvising without deep comprehension. This paper aims to address these doubts and provoke thoughts on the true nature of intelligence. Can an artificial intelligence system passing software engineering exams be deemed intelligent?\"}),/*#__PURE__*/e(\"p\",{children:\"GPT-4 exhibits sparks of artificial general intelligence through its core mental capabilities, range of expertise, and task versatility, but more work is needed to achieve complete AGI. The ultimate test is the ability to generate new knowledge, a task still beyond the capabilities of large language models.\"}),/*#__PURE__*/e(\"p\",{children:\"Nevertheless, evaluating the intelligence of large language models is necessary to ensure their reliability and effectiveness. A proper and comprehensive evaluation can detect errors, biases, and weaknesses, which can be utilized in improving their performance.\"}),/*#__PURE__*/e(\"h2\",{children:\"Learn more: We know how to measure the quality of LLMs\"}),/*#__PURE__*/t(\"p\",{children:[/*#__PURE__*/e(a,{href:\"https://toloka.ai/evaluation/\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:\"Toloka\u2019s Deep Evaluation platform\"})}),\" helps LLM developers evaluate their models effectively and produce better results. We achieve this by implementing customized quality metrics and human input to perform a thorough evaluation that matches your business needs.\"]}),/*#__PURE__*/e(\"h3\",{children:\"How does it work?\\xa0\"}),/*#__PURE__*/e(\"p\",{children:\"Our experts develop a custom evaluation plan:\"}),/*#__PURE__*/t(\"ol\",{children:[/*#__PURE__*/e(\"li\",{\"data-preset-tag\":\"p\",style:{\"--framer-font-size\":\"11px\",\"--framer-text-color\":\"rgb(0, 0, 0)\",\"--framer-text-decoration\":\"none\"},children:/*#__PURE__*/e(\"p\",{children:\"Review the usage scenarios and performance of the model\"})}),/*#__PURE__*/e(\"li\",{\"data-preset-tag\":\"p\",style:{\"--framer-font-size\":\"11px\",\"--framer-text-color\":\"rgb(0, 0, 0)\",\"--framer-text-decoration\":\"none\"},children:/*#__PURE__*/e(\"p\",{children:\"Formulate evaluation metrics tailored to your needs\"})}),/*#__PURE__*/e(\"li\",{\"data-preset-tag\":\"p\",style:{\"--framer-font-size\":\"11px\",\"--framer-text-color\":\"rgb(0, 0, 0)\",\"--framer-text-decoration\":\"none\"},children:/*#__PURE__*/e(\"p\",{children:\"Develop an evaluation pipeline with both automated and human annotation\"})}),/*#__PURE__*/e(\"li\",{\"data-preset-tag\":\"p\",style:{\"--framer-font-size\":\"11px\",\"--framer-text-color\":\"rgb(0, 0, 0)\",\"--framer-text-decoration\":\"none\"},children:/*#__PURE__*/e(\"p\",{children:\"Provide detailed reports for model improvement\"})})]}),/*#__PURE__*/t(\"p\",{children:[\"Want to leverage the full potential of your LLM? \",/*#__PURE__*/e(a,{href:\"https://toloka.ai/talk-to-us\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:\"Reach out\"})}),\" to discuss solutions.\"]})]});export const richText6=/*#__PURE__*/t(i.Fragment,{children:[/*#__PURE__*/e(\"p\",{children:\"The rapid advancement of large language models (LLMs) has led companies to increasingly turn to synthetic data\u2014AI-generated data that mimics real information. This technology marks the beginning of a new era, where the benefits of AI risk being overshadowed by the potential for harm. We believe that safe and responsible AI products require high-quality data grounded in human insight, with full transparency in data sourcing. While synthetic data is a powerful tool, and one we use ourselves, it still needs human insight for proper curation. Synthetic data is an important part of our offering, but not the only tool in our toolbox.\"}),/*#__PURE__*/e(\"p\",{children:\"Toloka\u2019s rich history puts our company in a unique position to embrace AI achievements and help shape AI development to better serve humanity.\"}),/*#__PURE__*/e(a,{href:\"https://toloka.ai/large-language-models/?utm_source=blog&utm_medium=banner&utm_campaign=sftrlhf&utm_content=evolutionoftoloka\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!0,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{className:\"framer-image\",\"data-preset-tag\":\"img\",children:/*#__PURE__*/e(\"img\",{alt:\"\",className:\"framer-image\",height:\"240\",src:\"https://framerusercontent.com/images/L1MWCdKVPqiR785bI54iY2TIc9k.png\",srcSet:\"https://framerusercontent.com/images/L1MWCdKVPqiR785bI54iY2TIc9k.png?scale-down-to=512 512w,https://framerusercontent.com/images/L1MWCdKVPqiR785bI54iY2TIc9k.png?scale-down-to=1024 1024w,https://framerusercontent.com/images/L1MWCdKVPqiR785bI54iY2TIc9k.png 1498w\",style:{aspectRatio:\"1498 / 480\"},width:\"749\"})})}),/*#__PURE__*/e(\"h2\",{children:\"Bridging Community and Technology\"}),/*#__PURE__*/e(\"p\",{children:\"There has always been a gap between AI technology and the data needed to fuel it, and Toloka was born out of a need to fill this gap with large-scale human data labeling. The project grew into one of the largest crowdsourcing platforms on the planet, with people all over the globe enriching data with a wealth of perspectives.\"}),/*#__PURE__*/e(\"p\",{children:\"The true strength of AI lies in its ability to reflect real-world experiences. Without human input, technology is useless to us; but without technology, human effort is not scalable. This understanding has guided our evolution over the past 10 years, continually seeking new ways to leverage the power of human knowledge and community with advanced technologies to harness it effectively.\"}),/*#__PURE__*/e(\"h2\",{children:\"The Role of Experts\"}),/*#__PURE__*/e(\"p\",{children:\"As large language models took the world by storm, the data labeling landscape changed dramatically. At first, human intuition and general knowledge were sufficient for training models. Now that foundation models perform well in general skills like answering basic questions, the game has changed. LLMs need refinement for enhanced performance in fields like coding, medicine, mathematics, automotive, finance \u2014 the list goes on. There is a high demand for specialized, dedicated datasets to solve complex tasks in specific domains.\"}),/*#__PURE__*/e(\"p\",{children:\"Toloka\u2019s focus has evolved from labeling existing data to crafting custom datasets from scratch, writing complex dialogs between AI agents and humans on niche topics. Instead of relying on aggregate knowledge from the crowd, we curate unique contributions from highly educated professionals like physicists, doctors, and software developers to craft specialized data samples to help train and improve LLMs.\"}),/*#__PURE__*/e(\"p\",{children:\"We recently established the Mindrift platform to bring together domain experts from around the world. The intention is to build on our experience scaling operational processes while grounding our efforts in cutting-edge research and our own insights as AI practitioners. The platform is designed to foster community in a deeper way, with experts in Mindrift focused on teamwork and collaboration to reach a goal, in stark contrast to the crowdsourcing model of anonymous contributions. Expert knowledge is funneled into data production pipelines that ensure high-quality results to support SFT, RLHF, and model evaluation. \"}),/*#__PURE__*/e(\"h2\",{children:\"A New Stage: Data Architecture\"}),/*#__PURE__*/e(\"p\",{children:\"Today, we stand at a pivotal moment where the role of a data architect is crucial. It's not enough to simply connect the dots between Tolokers with requesters; we must design efficient pipelines that augment human insight with synthetic data generated by LLMs. As architects, we analyze the model\u2019s needs and plan how to balance data types, and we run experiments to ensure the final dataset will truly make the model better.\"}),/*#__PURE__*/e(\"p\",{children:\"LLMs offer many opportunities to optimize data production, beyond generating synthetic data. We incorporate models into our pipelines with innovative co-pilots and auto checks that reduce routine work for our experts and improve datasets overall. We\u2019ve discovered a new level of synergy between human input and AI optimization.\"}),/*#__PURE__*/e(\"p\",{children:\"Toloka is building the foundation for future growth by refining our platform, expanding our community of professional experts, deepening our knowledge of the data we work with, and advancing our product to democratize access to this data for the AI community. \"})]});export const richText7=/*#__PURE__*/t(i.Fragment,{children:[/*#__PURE__*/e(\"p\",{children:\"We are pleased to announce our collaboration with NLP researchers and industry practitioners, including teams from MIT Lincoln Laboratory, Penn State University, and the University of Oslo, to develop a new benchmark for detecting artificial text.\"}),/*#__PURE__*/e(\"h3\",{children:\"Why is detecting AI-generated text so important?\"}),/*#__PURE__*/e(\"p\",{children:\"Detecting AI-generated text is crucial to prevent the spread of misinformation, maintain the quality of datasets used for training language models, and address ethical and legal concerns. Without detection, AI-generated content can be misrepresented as human-produced, leading to compromised data integrity and complex copyright issues. Moreover, reliance on AI-generated data can degrade the performance of models trained on such data.\"}),/*#__PURE__*/t(\"p\",{children:[\"Despite extensive research into detecting AI-generated content, many approaches fall short. Read our latest \",/*#__PURE__*/e(a,{href:\"https://toloka.ai/blog/ai-detection/\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:\"post\"})}),\" to learn more about current methods of AI detection and the challenges involved. \",/*#__PURE__*/e(\"br\",{}),/*#__PURE__*/e(\"br\",{className:\"trailing-break\"})]}),/*#__PURE__*/e(\"h3\",{children:\"Our approach to artificial text detector benchmarking\"}),/*#__PURE__*/e(\"p\",{children:\"Our benchmarking project differs from other approaches in that we compare LLM-generated texts to human-edited versions. We ask human annotators to edit AI-generated text by checking factuality, correcting grammar and awkward phrasing, removing hallucinations, and adding personal touches to make the text engaging and relevant. The two versions constitute a datapoint in the dataset: one representing an LLM-generated text, and the other a human-written version. The text examples cover a variety of use cases, from creative writing to summarizing news articles.\"}),/*#__PURE__*/e(\"p\",{children:\"The resulting dataset can be used to build and test new artificial text detectors.\"}),/*#__PURE__*/e(\"h3\",{children:\"Get involved\"}),/*#__PURE__*/t(\"p\",{children:[\"We invite you to contribute to the dataset and make a difference! To get started, click the Start button on the \",/*#__PURE__*/e(a,{href:\"https://toloka.ai/ai-text-detection#how-can-i-contribute\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:\"project page\"})}),\" and do a trial task.\"]}),/*#__PURE__*/e(\"h3\",{children:\"Get early access to benchmarks \"}),/*#__PURE__*/e(\"p\",{children:\"The open dataset will be published in September. Get on the mailing list now to be the first to know when benchmarks are available!\"}),/*#__PURE__*/e(\"p\",{children:/*#__PURE__*/e(a,{href:\"https://toloka.ai/ai-text-detection\",motionChild:!0,nodeId:\"eo4RAmtig\",openInNewTab:!1,scopeId:\"contentManagement\",smoothScroll:!1,children:/*#__PURE__*/e(n.a,{children:\"Subscribe to updates\"})})})]});\nexport const __FramerMetadata__ = {\"exports\":{\"richText\":{\"type\":\"variable\",\"annotations\":{\"framerContractVersion\":\"1\"}},\"richText2\":{\"type\":\"variable\",\"annotations\":{\"framerContractVersion\":\"1\"}},\"richText1\":{\"type\":\"variable\",\"annotations\":{\"framerContractVersion\":\"1\"}},\"richText3\":{\"type\":\"variable\",\"annotations\":{\"framerContractVersion\":\"1\"}},\"richText4\":{\"type\":\"variable\",\"annotations\":{\"framerContractVersion\":\"1\"}},\"richText5\":{\"type\":\"variable\",\"annotations\":{\"framerContractVersion\":\"1\"}},\"richText7\":{\"type\":\"variable\",\"annotations\":{\"framerContractVersion\":\"1\"}},\"richText6\":{\"type\":\"variable\",\"annotations\":{\"framerContractVersion\":\"1\"}},\"__FramerMetadata__\":{\"type\":\"variable\"}}}"],
  "mappings": "6JAAsJ,IAAMA,EAAsBC,EAAIC,EAAS,CAAC,SAAS,CAAcC,EAAE,KAAK,CAAC,SAAsBA,EAAE,KAAK,CAAC,UAAU,gBAAgB,CAAC,CAAC,CAAC,EAAeF,EAAE,KAAK,CAAC,SAAS,CAAC,UAAuBE,EAAEC,EAAE,CAAC,KAAK,qHAAqH,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAS,gBAAgB,CAAC,CAAC,CAAC,EAAE,0FAA0F,CAAC,CAAC,EAAeF,EAAEC,EAAE,CAAC,KAAK,+IAA+I,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,UAAU,eAAe,kBAAkB,MAAM,SAAsBF,EAAE,MAAM,CAAC,IAAI,GAAG,UAAU,eAAe,OAAO,MAAM,IAAI,uEAAuE,OAAO,uQAAuQ,MAAM,CAAC,YAAY,YAAY,EAAE,MAAM,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,gVAAgV,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,mWAAmW,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,qUAAqU,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,wBAAwB,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,6SAA6S,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,oUAAoU,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,gZAAgZ,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,kQAAkQ,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,iOAAiO,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,yDAAyD,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,yYAAyY,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,yCAAyC,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,iRAAiR,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,iSAAiS,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,6aAA6a,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,0BAA0B,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,mDAAmD,CAAC,EAAeF,EAAE,KAAK,CAAC,SAAS,CAAcE,EAAEC,EAAE,CAAC,KAAK,qHAAqH,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAS,2BAA2B,CAAC,CAAC,CAAC,EAAE,yEAAyE,CAAC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,sNAAsN,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,wOAAmO,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,+BAA+B,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,4UAAuU,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,2BAA2B,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,8QAA8Q,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,yBAAyB,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,iSAAiS,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,wBAAwB,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,0dAA0d,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,8dAA8d,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,oCAAoC,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,4QAAuQ,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,2TAA2T,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,uhBAAuhB,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,0SAA0S,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,oSAAoS,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,iiBAA4hB,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,ocAAoc,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,sPAAsP,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,+UAA+U,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,sCAAsC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAC,sEAAmFE,EAAEC,EAAE,CAAC,KAAK,mCAAmC,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAS,kDAAkD,CAAC,CAAC,CAAC,EAAE,wWAAwW,CAAC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,6jBAA6jB,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,oXAA+W,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,gVAAgV,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,sbAA4a,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,+YAA+Y,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,obAA+a,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,kdAAkd,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,ufAAuf,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,qCAAqC,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,kRAAkR,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAC,sLAAmME,EAAEC,EAAE,CAAC,KAAK,mIAAmI,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAS,oDAAoD,CAAC,CAAC,CAAC,EAAE,yIAAyI,CAAC,CAAC,EAAeF,EAAE,KAAK,CAAC,SAAS,iBAAiB,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAC,iGAA8GE,EAAEC,EAAE,CAAC,KAAK,qEAAqE,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAS,yBAAyB,CAAC,CAAC,CAAC,EAAE,6VAA6V,CAAC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,oTAA+S,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,sYAAiY,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,6cAA6c,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,gXAA2W,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,2aAAsa,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,oCAAoC,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,6TAAwT,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,wUAAwU,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAC,sBAAmCE,EAAEC,EAAE,CAAC,KAAK,mCAAmC,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAS,yDAAyD,CAAC,CAAC,CAAC,EAAE,0PAA0P,CAAC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,4NAA4N,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,cAAc,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,+WAA+W,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,4cAAuc,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,uTAAkT,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,ygBAAygB,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,yRAAyR,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,uRAAuR,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAC,oBAAiCE,EAAEC,EAAE,CAAC,KAAK,sCAAsC,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAS,+DAA+D,CAAC,CAAC,CAAC,EAAE,yIAAyI,CAAC,CAAC,EAAeF,EAAE,KAAK,CAAC,SAAS,eAAe,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,4BAAuB,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAcE,EAAE,SAAS,CAAC,SAAS,sBAAsB,CAAC,EAAE,4RAA4R,CAAC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAcE,EAAE,SAAS,CAAC,SAAS,sBAAsB,CAAC,EAAE,4dAAkd,CAAC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAcE,EAAE,SAAS,CAAC,SAAS,qBAAqB,CAAC,EAAE,qSAAqS,CAAC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAcE,EAAE,SAAS,CAAC,SAAS,gCAAgC,CAAC,EAAE,8SAAyS,CAAC,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,oeAAoe,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,oCAAoC,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,mBAAmB,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,wTAAwT,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,6BAA6B,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,kVAAkV,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,4OAA4O,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,0BAA0B,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,mLAAmL,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,qBAAqB,CAAC,EAAeA,EAAE,MAAM,CAAC,IAAI,GAAG,UAAU,eAAe,OAAO,MAAM,IAAI,uEAAuE,OAAO,uQAAuQ,MAAM,CAAC,YAAY,aAAa,EAAE,MAAM,KAAK,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,OAAO,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,gaAAga,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,oPAAoP,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,gVAAgV,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,ieAAie,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,+CAA+C,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,qZAAqZ,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAC,gBAA6BE,EAAEC,EAAE,CAAC,KAAK,mCAAmC,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAsBF,EAAE,SAAS,CAAC,SAAS,oDAAoD,CAAC,CAAC,CAAC,CAAC,CAAC,EAAE,0ZAA0Z,CAAC,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,sBAAsB,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,oBAAoB,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,uWAAuW,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,oBAAoB,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,idAAid,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,wBAAwB,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,iTAAiT,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,yBAAyB,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,iWAAiW,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,+BAA+B,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,gRAA2Q,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,6BAA6B,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,2FAA2F,CAAC,EAAeF,EAAE,KAAK,CAAC,SAAS,CAAcE,EAAE,KAAK,CAAC,kBAAkB,IAAI,MAAM,CAAC,qBAAqB,OAAO,sBAAsB,eAAe,2BAA2B,MAAM,EAAE,SAAsBF,EAAE,IAAI,CAAC,SAAS,CAAcE,EAAE,SAAS,CAAC,SAAS,kBAAkB,CAAC,EAAE,sOAAsO,CAAC,CAAC,CAAC,CAAC,EAAeA,EAAE,KAAK,CAAC,kBAAkB,IAAI,MAAM,CAAC,qBAAqB,OAAO,sBAAsB,eAAe,2BAA2B,MAAM,EAAE,SAAsBF,EAAE,IAAI,CAAC,SAAS,CAAcE,EAAE,SAAS,CAAC,SAAS,kBAAkB,CAAC,EAAE,kQAAkQ,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,qBAAqB,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,oWAAoW,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,iCAAiC,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,oeAAoe,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,2BAA2B,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,oRAAoR,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,yCAAyC,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,mXAAmX,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,sXAAsX,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,yDAAyD,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,8DAA8D,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,wcAAmc,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,4gBAA4gB,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,sDAAsD,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,oYAA+X,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,sfAAif,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,qDAAqD,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,saAAia,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,kXAA6W,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,0RAA0R,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,+BAA+B,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,0OAA0O,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,0aAA0a,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,ocAAoc,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,kfAAkf,CAAC,CAAC,CAAC,CAAC,EAAeG,EAAuBL,EAAIC,EAAS,CAAC,SAAS,CAAcC,EAAE,IAAI,CAAC,SAAS,uUAAuU,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,+VAA0V,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,yRAAyR,CAAC,EAAeA,EAAE,MAAM,CAAC,IAAI,+DAA+D,UAAU,eAAe,OAAO,MAAM,IAAI,uEAAuE,MAAM,CAAC,YAAY,WAAW,EAAE,MAAM,KAAK,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAcE,EAAE,KAAK,CAAC,SAAS,wEAAwE,CAAC,EAAeA,EAAEC,EAAE,CAAC,KAAK,4FAA4F,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAsBF,EAAE,KAAK,CAAC,SAAS,gDAA2C,CAAC,CAAC,CAAC,CAAC,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,GAAG,CAAC,CAAC,CAAC,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,mTAAmT,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,kPAA6O,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,iCAAiC,CAAC,EAAeF,EAAE,KAAK,CAAC,SAAS,CAAC,gJAA6JE,EAAEC,EAAE,CAAC,KAAK,0GAA0G,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAS,OAAO,CAAC,CAAC,CAAC,EAAE,GAAG,CAAC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,icAAic,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,iXAAuW,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,sQAAsQ,CAAC,EAAeA,EAAE,MAAM,CAAC,IAAI,mFAAmF,UAAU,eAAe,OAAO,MAAM,IAAI,uEAAuE,MAAM,CAAC,YAAY,WAAW,EAAE,MAAM,KAAK,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAcE,EAAE,KAAK,CAAC,SAAS,qIAAqI,CAAC,EAAeA,EAAEC,EAAE,CAAC,KAAK,mCAAmC,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAsBF,EAAE,KAAK,CAAC,SAAS,gEAAgE,CAAC,CAAC,CAAC,CAAC,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,GAAG,CAAC,CAAC,CAAC,CAAC,EAAeA,EAAEC,EAAE,CAAC,KAAK,gJAAgJ,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,UAAU,eAAe,kBAAkB,MAAM,SAAsBF,EAAE,MAAM,CAAC,IAAI,GAAG,UAAU,eAAe,OAAO,MAAM,IAAI,uEAAuE,OAAO,uQAAuQ,MAAM,CAAC,YAAY,YAAY,EAAE,MAAM,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,4CAA4C,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,mYAAmY,CAAC,EAAeA,EAAE,MAAM,CAAC,IAAI,GAAG,UAAU,eAAe,OAAO,MAAM,IAAI,sEAAsE,OAAO,sKAAsK,MAAM,CAAC,YAAY,WAAW,EAAE,MAAM,KAAK,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAcE,EAAE,KAAK,CAAC,SAAS,yLAAyL,CAAC,EAAeA,EAAEC,EAAE,CAAC,KAAK,oJAAoJ,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAsBF,EAAE,KAAK,CAAC,SAAS,WAAW,CAAC,CAAC,CAAC,CAAC,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,GAAG,CAAC,CAAC,CAAC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAC,wBAAqCE,EAAEC,EAAE,CAAC,KAAK,sEAAsE,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAS,YAAY,CAAC,CAAC,CAAC,EAAE,4VAA4V,CAAC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,gSAAsR,CAAC,EAAeA,EAAE,MAAM,CAAC,IAAI,GAAG,UAAU,eAAe,OAAO,MAAM,IAAI,uEAAuE,OAAO,wKAAwK,MAAM,CAAC,YAAY,WAAW,EAAE,MAAM,KAAK,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAcE,EAAE,KAAK,CAAC,SAAS,mIAAmI,CAAC,EAAeA,EAAEC,EAAE,CAAC,KAAK,6DAA6D,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAsBF,EAAE,KAAK,CAAC,SAAS,sDAAsD,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAC,iNAA8NE,EAAEC,EAAE,CAAC,KAAK,kPAAkP,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAS,4EAA4E,CAAC,CAAC,CAAC,EAAE,8JAA8J,CAAC,CAAC,EAAeF,EAAE,KAAK,CAAC,SAAS,wDAAwD,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,2GAA2G,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,8CAA8C,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,mPAAmP,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,+DAA+D,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,4PAA4P,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,4DAA4D,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,oSAAoS,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,uBAAuB,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,qOAAqO,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,kRAAkR,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,0BAA0B,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,2OAAsO,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,qBAAqB,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,8RAA8R,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,qPAAqP,CAAC,EAAeA,EAAE,MAAM,CAAC,IAAI,GAAG,UAAU,eAAe,OAAO,MAAM,IAAI,sEAAsE,MAAM,CAAC,YAAY,WAAW,EAAE,MAAM,KAAK,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAcE,EAAE,KAAK,CAAC,SAAS,mHAAmH,CAAC,EAAeA,EAAEC,EAAE,CAAC,KAAK,0CAA0C,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAsBF,EAAE,KAAK,CAAC,SAAS,eAAe,CAAC,CAAC,CAAC,CAAC,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,GAAG,CAAC,CAAC,CAAC,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,wCAAwC,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,iQAAiQ,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,8PAA8P,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,iYAAiY,CAAC,EAAeA,EAAE,MAAM,CAAC,IAAI,GAAG,UAAU,eAAe,OAAO,MAAM,IAAI,uEAAuE,OAAO,4EAA4E,MAAM,CAAC,YAAY,WAAW,EAAE,MAAM,KAAK,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAcE,EAAE,KAAK,CAAC,SAAS,gKAAgK,CAAC,EAAeA,EAAEC,EAAE,CAAC,KAAK,mCAAmC,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAsBF,EAAE,KAAK,CAAC,SAAS,mDAAmD,CAAC,CAAC,CAAC,CAAC,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,GAAG,CAAC,CAAC,CAAC,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,qOAAqO,CAAC,EAAeA,EAAE,MAAM,CAAC,IAAI,GAAG,UAAU,eAAe,OAAO,MAAM,IAAI,uEAAuE,OAAO,wKAAwK,MAAM,CAAC,YAAY,WAAW,EAAE,MAAM,KAAK,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAcE,EAAE,KAAK,CAAC,SAAS,iIAAiI,CAAC,EAAeA,EAAEC,EAAE,CAAC,KAAK,mCAAmC,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAsBF,EAAE,KAAK,CAAC,SAAS,wEAAwE,CAAC,CAAC,CAAC,CAAC,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,GAAG,CAAC,CAAC,CAAC,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,6BAA6B,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,oMAAoM,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAC,gBAA6BE,EAAEC,EAAE,CAAC,KAAK,4CAA4C,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAS,QAAQ,CAAC,CAAC,CAAC,EAAE,iTAAiT,CAAC,CAAC,EAAeF,EAAE,MAAM,CAAC,IAAI,GAAG,UAAU,eAAe,OAAO,MAAM,IAAI,qEAAqE,OAAO,qKAAqK,MAAM,CAAC,YAAY,YAAY,EAAE,MAAM,KAAK,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAcE,EAAE,KAAK,CAAC,SAAS,iNAAiN,CAAC,EAAeA,EAAEC,EAAE,CAAC,KAAK,4CAA4C,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAsBF,EAAE,KAAK,CAAC,SAAS,cAAc,CAAC,CAAC,CAAC,CAAC,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,GAAG,CAAC,CAAC,CAAC,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,2LAA2L,CAAC,EAAeA,EAAE,MAAM,CAAC,IAAI,GAAG,UAAU,eAAe,OAAO,MAAM,IAAI,qEAAqE,OAAO,iQAAiQ,MAAM,CAAC,YAAY,YAAY,EAAE,MAAM,KAAK,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAcE,EAAE,KAAK,CAAC,SAAS,8JAA8J,CAAC,EAAeA,EAAEC,EAAE,CAAC,KAAK,mCAAmC,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAsBF,EAAE,KAAK,CAAC,SAAS,8FAA8F,CAAC,CAAC,CAAC,CAAC,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,GAAG,CAAC,CAAC,CAAC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAcE,EAAEC,EAAE,CAAC,KAAK,mCAAmC,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAS,qBAAqB,CAAC,CAAC,CAAC,EAAE,oMAAoM,CAAC,CAAC,EAAeF,EAAE,MAAM,CAAC,IAAI,GAAG,UAAU,eAAe,OAAO,MAAM,IAAI,sEAAsE,OAAO,sKAAsK,MAAM,CAAC,YAAY,WAAW,EAAE,MAAM,KAAK,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAcE,EAAE,KAAK,CAAC,SAAS,gIAAgI,CAAC,EAAeA,EAAEC,EAAE,CAAC,KAAK,mCAAmC,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAsBF,EAAE,KAAK,CAAC,SAAS,sBAAsB,CAAC,CAAC,CAAC,CAAC,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,GAAG,CAAC,CAAC,CAAC,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,4QAA4Q,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,yCAAyC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAC,yjBAAskBE,EAAE,KAAK,CAAC,SAAS,QAAQ,CAAC,EAAeA,EAAEC,EAAE,CAAC,KAAK,iDAAiD,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAsBF,EAAE,KAAK,CAAC,SAAS,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,yBAAyB,CAAC,CAAC,CAAC,CAAC,EAAeF,EAAE,KAAK,CAAC,SAAS,CAAC,kJAA0JE,EAAEC,EAAE,CAAC,KAAK,0GAA0G,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAS,YAAY,CAAC,CAAC,CAAC,EAAE,cAAc,CAAC,CAAC,CAAC,CAAC,CAAC,EAAeE,EAAuBN,EAAIC,EAAS,CAAC,SAAS,CAAcC,EAAE,IAAI,CAAC,SAAS,iMAAiM,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,kRAAkR,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,2KAAsK,CAAC,EAAeA,EAAEC,EAAE,CAAC,KAAK,wHAAwH,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,UAAU,eAAe,kBAAkB,MAAM,SAAsBF,EAAE,MAAM,CAAC,IAAI,GAAG,UAAU,eAAe,OAAO,MAAM,IAAI,sEAAsE,OAAO,oQAAoQ,MAAM,CAAC,YAAY,YAAY,EAAE,MAAM,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,4CAA4C,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,8dAA+c,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,kcAAkc,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,kCAAkC,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,2hBAAshB,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAC,wRAA2RE,EAAEC,EAAE,CAAC,KAAK,yCAAyC,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAS,eAAe,CAAC,CAAC,CAAC,EAAE,wIAAwI,CAAC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,mTAAmT,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,gCAAgC,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,igBAAigB,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,+cAA+c,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,0WAA0W,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,+OAA0O,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,mCAAmC,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,mSAA8R,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAC,oQAAiRE,EAAEC,EAAE,CAAC,KAAK,uCAAuC,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAS,8BAA8B,CAAC,CAAC,CAAC,EAAE,GAAG,CAAC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,gMAAgM,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,iDAAiD,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,iLAA4K,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,0MAA0M,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,iZAAiZ,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,gVAAsU,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,2BAA2B,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,uNAAuN,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAC,4BAAyCE,EAAEC,EAAE,CAAC,KAAK,4EAA4E,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAS,uCAAuC,CAAC,CAAC,CAAC,EAAE,iIAAiI,CAAC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,wOAAwO,CAAC,EAAeF,EAAE,KAAK,CAAC,SAAS,CAAcE,EAAE,KAAK,CAAC,kBAAkB,IAAI,MAAM,CAAC,qBAAqB,OAAO,sBAAsB,eAAe,2BAA2B,MAAM,EAAE,SAAsBA,EAAE,IAAI,CAAC,SAAS,qGAAqG,CAAC,CAAC,CAAC,EAAeA,EAAE,KAAK,CAAC,kBAAkB,IAAI,MAAM,CAAC,qBAAqB,OAAO,sBAAsB,eAAe,2BAA2B,MAAM,EAAE,SAAsBA,EAAE,IAAI,CAAC,SAAS,mEAAmE,CAAC,CAAC,CAAC,EAAeA,EAAE,KAAK,CAAC,kBAAkB,IAAI,MAAM,CAAC,qBAAqB,OAAO,sBAAsB,eAAe,2BAA2B,MAAM,EAAE,SAAsBA,EAAE,IAAI,CAAC,SAAS,6GAA6G,CAAC,CAAC,CAAC,EAAeA,EAAE,KAAK,CAAC,kBAAkB,IAAI,MAAM,CAAC,qBAAqB,OAAO,sBAAsB,eAAe,2BAA2B,MAAM,EAAE,SAAsBA,EAAE,IAAI,CAAC,SAAS,mCAAmC,CAAC,CAAC,CAAC,EAAeA,EAAE,KAAK,CAAC,kBAAkB,IAAI,MAAM,CAAC,qBAAqB,OAAO,sBAAsB,eAAe,2BAA2B,MAAM,EAAE,SAAsBA,EAAE,IAAI,CAAC,SAAS,iCAAiC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,mYAAoX,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAcE,EAAEC,EAAE,CAAC,KAAK,+BAA+B,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAS,gCAAgC,CAAC,CAAC,CAAC,EAAE,YAAY,CAAC,CAAC,CAAC,CAAC,CAAC,EAAeG,EAAuBP,EAAIC,EAAS,CAAC,SAAS,CAAcC,EAAE,IAAI,CAAC,SAAS,gTAAgT,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,yYAAoY,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,sBAAsB,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,saAAsa,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,odAA0c,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,uUAAuU,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,kBAAkB,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAC,4JAAyKE,EAAEC,EAAE,CAAC,KAAK,qIAAqI,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAS,iBAAiB,CAAC,CAAC,CAAC,EAAE,wfAAwf,CAAC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,6TAA6T,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,kOAAkO,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,sVAAsV,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,sBAAsB,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,sSAAsS,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,sfAAsf,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,uRAAuR,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,iCAAiC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAcE,EAAE,SAAS,CAAC,SAAS,mBAAmB,CAAC,EAAE,iQAAiQ,CAAC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAcE,EAAE,SAAS,CAAC,SAAS,cAAc,CAAC,EAAE,sQAAsQ,CAAC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAcE,EAAE,SAAS,CAAC,SAAS,gBAAgB,CAAC,EAAE,gNAAgN,CAAC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAcE,EAAE,SAAS,CAAC,SAAS,mBAAmB,CAAC,EAAE,kPAAkP,CAAC,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,iCAAiC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAcE,EAAE,SAAS,CAAC,SAAS,qCAAqC,CAAC,EAAE,4KAA4K,CAAC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAcE,EAAE,SAAS,CAAC,SAAS,gBAAgB,CAAC,EAAE,wLAAwL,CAAC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAcE,EAAE,SAAS,CAAC,SAAS,gBAAgB,CAAC,EAAE,uOAAuO,CAAC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAcE,EAAE,SAAS,CAAC,SAAS,sCAAsC,CAAC,EAAE,iJAAiJ,CAAC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAcE,EAAE,SAAS,CAAC,SAAS,yBAAyB,CAAC,EAAE,6OAA6O,CAAC,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,sBAAsB,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,8fAA8f,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,6cAA6c,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,kZAA6Y,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,6SAA6S,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,geAAge,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,6CAA6C,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,miBAAmiB,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,+mBAA+mB,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,mgBAAmgB,CAAC,CAAC,CAAC,CAAC,EAAeM,EAAuBR,EAAIC,EAAS,CAAC,SAAS,CAAcC,EAAE,IAAI,CAAC,SAAS,icAAic,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,2BAA2B,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,kVAAkV,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,mYAAmY,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,qUAAqU,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,uUAAuU,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,uWAAuW,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,0BAA0B,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,uBAAuB,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,qcAAqc,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,0cAA0c,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,uBAAuB,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,uNAAuN,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,4bAA4b,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,0kBAA0kB,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,iaAAia,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,wZAAwZ,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,4RAA4R,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,sEAAsE,CAAC,EAAeA,EAAE,MAAM,CAAC,IAAI,GAAG,UAAU,eAAe,OAAO,MAAM,IAAI,sEAAsE,OAAO,oQAAoQ,MAAM,CAAC,YAAY,YAAY,EAAE,MAAM,KAAK,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,yBAAyB,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,mYAAmY,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,uXAAuX,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,mCAAmC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAcE,EAAE,SAAS,CAAC,SAAS,kCAAkC,CAAC,EAAE,8OAA8O,CAAC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAcE,EAAE,SAAS,CAAC,SAAS,sCAAsC,CAAC,EAAE,qOAAqO,CAAC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAcE,EAAE,SAAS,CAAC,SAAS,SAAS,CAAC,EAAE,sdAAsd,CAAC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAcE,EAAE,SAAS,CAAC,SAAS,iDAAiD,CAAC,EAAE,+eAA+e,CAAC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAcE,EAAE,SAAS,CAAC,SAAS,sBAAsB,CAAC,EAAE,oTAAoT,CAAC,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,iCAAiC,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,+BAA+B,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,4RAA4R,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,gDAAgD,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,wQAAwQ,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,uCAAuC,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,wRAAwR,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,sCAAsC,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,8QAA8Q,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,4BAA4B,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAsBA,EAAE,SAAS,CAAC,SAAS,SAAS,CAAC,CAAC,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,+fAA+f,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,6gBAAwgB,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAsBA,EAAE,SAAS,CAAC,SAAS,OAAO,CAAC,CAAC,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,ieAAie,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,kCAAkC,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,8TAA8T,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,kdAAkd,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,gRAAgR,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,iCAAiC,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,ySAAyS,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,4NAA4N,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,gTAAgT,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,qSAAqS,CAAC,CAAC,CAAC,CAAC,EAAeO,EAAuBT,EAAIC,EAAS,CAAC,SAAS,CAAcD,EAAE,IAAI,CAAC,SAAS,CAAC,yIAAsJE,EAAEC,EAAE,CAAC,KAAK,mCAAmC,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAS,gBAAgB,CAAC,CAAC,CAAC,EAAE,ozBAAszB,CAAC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,kHAAkH,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,6CAA6C,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,u2BAAu2B,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,sRAAsR,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,wDAAwD,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,mRAAmR,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,mDAAmD,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,iOAAiO,CAAC,EAAeA,EAAE,MAAM,CAAC,IAAI,GAAG,UAAU,eAAe,OAAO,MAAM,IAAI,uEAAuE,OAAO,wKAAwK,MAAM,CAAC,YAAY,WAAW,EAAE,MAAM,KAAK,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAC,OAAoBE,EAAEC,EAAE,CAAC,KAAK,mCAAmC,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAS,4CAA4C,CAAC,CAAC,CAAC,EAAE,yBAAsCF,EAAE,KAAK,CAAC,CAAC,EAAeA,EAAE,KAAK,CAAC,UAAU,gBAAgB,CAAC,CAAC,CAAC,CAAC,EAAeA,EAAE,MAAM,CAAC,IAAI,GAAG,UAAU,eAAe,OAAO,MAAM,IAAI,qEAAqE,OAAO,oKAAoK,MAAM,CAAC,YAAY,WAAW,EAAE,MAAM,KAAK,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAC,OAAoBE,EAAEC,EAAE,CAAC,KAAK,mCAAmC,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAS,4CAA4C,CAAC,CAAC,CAAC,EAAE,wBAAwB,CAAC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,4ZAA4Z,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,gCAAgC,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,sjBAAsjB,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,+BAA+B,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,uYAAuY,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,kYAAkY,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,mOAAmO,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,2lBAA2lB,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,0CAA0C,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,kMAAkM,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,yZAAyZ,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,mTAAmT,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,8PAA8P,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,kDAAkD,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,ooBAAooB,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,mbAAmb,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,mDAAmD,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,8RAA8R,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,4HAA4H,CAAC,EAAeA,EAAE,MAAM,CAAC,IAAI,GAAG,UAAU,eAAe,OAAO,MAAM,IAAI,sEAAsE,OAAO,oQAAoQ,MAAM,CAAC,YAAY,YAAY,EAAE,MAAM,KAAK,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAC,OAAoBE,EAAEC,EAAE,CAAC,KAAK,mCAAmC,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAS,4CAA4C,CAAC,CAAC,CAAC,EAAE,wBAAwB,CAAC,CAAC,EAAeF,EAAE,KAAK,CAAC,SAAS,4CAA4C,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,oHAAoH,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,uzBAAuzB,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,yIAAyI,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,gSAAgS,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,qMAAqM,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,qFAAqF,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,ubAAub,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,sTAAsT,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,uQAAuQ,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,wDAAwD,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAcE,EAAEC,EAAE,CAAC,KAAK,gCAAgC,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAS,wCAAmC,CAAC,CAAC,CAAC,EAAE,mOAAmO,CAAC,CAAC,EAAeF,EAAE,KAAK,CAAC,SAAS,uBAAuB,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,+CAA+C,CAAC,EAAeF,EAAE,KAAK,CAAC,SAAS,CAAcE,EAAE,KAAK,CAAC,kBAAkB,IAAI,MAAM,CAAC,qBAAqB,OAAO,sBAAsB,eAAe,2BAA2B,MAAM,EAAE,SAAsBA,EAAE,IAAI,CAAC,SAAS,yDAAyD,CAAC,CAAC,CAAC,EAAeA,EAAE,KAAK,CAAC,kBAAkB,IAAI,MAAM,CAAC,qBAAqB,OAAO,sBAAsB,eAAe,2BAA2B,MAAM,EAAE,SAAsBA,EAAE,IAAI,CAAC,SAAS,qDAAqD,CAAC,CAAC,CAAC,EAAeA,EAAE,KAAK,CAAC,kBAAkB,IAAI,MAAM,CAAC,qBAAqB,OAAO,sBAAsB,eAAe,2BAA2B,MAAM,EAAE,SAAsBA,EAAE,IAAI,CAAC,SAAS,yEAAyE,CAAC,CAAC,CAAC,EAAeA,EAAE,KAAK,CAAC,kBAAkB,IAAI,MAAM,CAAC,qBAAqB,OAAO,sBAAsB,eAAe,2BAA2B,MAAM,EAAE,SAAsBA,EAAE,IAAI,CAAC,SAAS,gDAAgD,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAC,oDAAiEE,EAAEC,EAAE,CAAC,KAAK,+BAA+B,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAS,WAAW,CAAC,CAAC,CAAC,EAAE,wBAAwB,CAAC,CAAC,CAAC,CAAC,CAAC,EAAeM,EAAuBV,EAAIC,EAAS,CAAC,SAAS,CAAcC,EAAE,IAAI,CAAC,SAAS,koBAA6nB,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,qJAAgJ,CAAC,EAAeA,EAAEC,EAAE,CAAC,KAAK,gIAAgI,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,UAAU,eAAe,kBAAkB,MAAM,SAAsBF,EAAE,MAAM,CAAC,IAAI,GAAG,UAAU,eAAe,OAAO,MAAM,IAAI,uEAAuE,OAAO,uQAAuQ,MAAM,CAAC,YAAY,YAAY,EAAE,MAAM,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,mCAAmC,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,yUAAyU,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,sYAAsY,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,qBAAqB,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,0hBAAqhB,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,6ZAAwZ,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,inBAAinB,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,gCAAgC,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,gbAA2a,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,8UAAyU,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,sQAAsQ,CAAC,CAAC,CAAC,CAAC,EAAeS,EAAuBX,EAAIC,EAAS,CAAC,SAAS,CAAcC,EAAE,IAAI,CAAC,SAAS,yPAAyP,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,kDAAkD,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,sbAAsb,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAC,+GAA4HE,EAAEC,EAAE,CAAC,KAAK,uCAAuC,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAS,MAAM,CAAC,CAAC,CAAC,EAAE,qFAAkGF,EAAE,KAAK,CAAC,CAAC,EAAeA,EAAE,KAAK,CAAC,UAAU,gBAAgB,CAAC,CAAC,CAAC,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,uDAAuD,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,ojBAAojB,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,oFAAoF,CAAC,EAAeA,EAAE,KAAK,CAAC,SAAS,cAAc,CAAC,EAAeF,EAAE,IAAI,CAAC,SAAS,CAAC,mHAAgIE,EAAEC,EAAE,CAAC,KAAK,2DAA2D,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAS,cAAc,CAAC,CAAC,CAAC,EAAE,uBAAuB,CAAC,CAAC,EAAeF,EAAE,KAAK,CAAC,SAAS,iCAAiC,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAS,qIAAqI,CAAC,EAAeA,EAAE,IAAI,CAAC,SAAsBA,EAAEC,EAAE,CAAC,KAAK,sCAAsC,YAAY,GAAG,OAAO,YAAY,aAAa,GAAG,QAAQ,oBAAoB,aAAa,GAAG,SAAsBD,EAAEE,EAAE,EAAE,CAAC,SAAS,sBAAsB,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,EAC9kiIQ,EAAqB,CAAC,QAAU,CAAC,SAAW,CAAC,KAAO,WAAW,YAAc,CAAC,sBAAwB,GAAG,CAAC,EAAE,UAAY,CAAC,KAAO,WAAW,YAAc,CAAC,sBAAwB,GAAG,CAAC,EAAE,UAAY,CAAC,KAAO,WAAW,YAAc,CAAC,sBAAwB,GAAG,CAAC,EAAE,UAAY,CAAC,KAAO,WAAW,YAAc,CAAC,sBAAwB,GAAG,CAAC,EAAE,UAAY,CAAC,KAAO,WAAW,YAAc,CAAC,sBAAwB,GAAG,CAAC,EAAE,UAAY,CAAC,KAAO,WAAW,YAAc,CAAC,sBAAwB,GAAG,CAAC,EAAE,UAAY,CAAC,KAAO,WAAW,YAAc,CAAC,sBAAwB,GAAG,CAAC,EAAE,UAAY,CAAC,KAAO,WAAW,YAAc,CAAC,sBAAwB,GAAG,CAAC,EAAE,mBAAqB,CAAC,KAAO,UAAU,CAAC,CAAC",
  "names": ["richText", "u", "x", "p", "Link", "motion", "richText1", "richText2", "richText3", "richText4", "richText5", "richText6", "richText7", "__FramerMetadata__"]
}