deploy: b4cc0126df
This commit is contained in:
eric
@@ -1,7 +1,7 @@
|
||||
<!doctype html><html lang=en><head><title>A Deep Dive into PPO for Language Models · Eric X. Liu's Personal Page</title><meta charset=utf-8><meta name=viewport content="width=device-width,initial-scale=1"><meta name=color-scheme content="light dark"><meta name=author content="Eric X. Liu"><meta name=description content="Large Language Models (LLMs) have demonstrated astonishing capabilities, but out-of-the-box, they are simply powerful text predictors. They don’t inherently understand what makes a response helpful, harmless, or aligned with human values. The technique that has proven most effective at bridging this gap is Reinforcement Learning from Human Feedback (RLHF), and at its heart lies a powerful algorithm: Proximal Policy Optimization (PPO).
|
||||
You may have seen diagrams like the one below, which outlines the RLHF training process. It can look intimidating, with a web of interconnected models, losses, and data flows."><meta name=keywords content="software engineer,performance engineering,Google engineer,tech blog,software development,performance optimization,Eric Liu,engineering blog,mountain biking,Jeep enthusiast,overlanding,camping,outdoor adventures"><meta name=fediverse:creator content><meta name=twitter:card content="summary"><meta name=twitter:title content="A Deep Dive into PPO for Language Models"><meta name=twitter:description content="Large Language Models (LLMs) have demonstrated astonishing capabilities, but out-of-the-box, they are simply powerful text predictors. They don’t inherently understand what makes a response helpful, harmless, or aligned with human values. The technique that has proven most effective at bridging this gap is Reinforcement Learning from Human Feedback (RLHF), and at its heart lies a powerful algorithm: Proximal Policy Optimization (PPO).
|
||||
You may have seen diagrams like the one below, which outlines the RLHF training process. It can look intimidating, with a web of interconnected models, losses, and data flows."><meta name=keywords content="software engineer,performance engineering,Google engineer,tech blog,software development,performance optimization,Eric Liu,engineering blog,mountain biking,Jeep enthusiast,overlanding,camping,outdoor adventures"><meta name=twitter:card content="summary"><meta name=twitter:title content="A Deep Dive into PPO for Language Models"><meta name=twitter:description content="Large Language Models (LLMs) have demonstrated astonishing capabilities, but out-of-the-box, they are simply powerful text predictors. They don’t inherently understand what makes a response helpful, harmless, or aligned with human values. The technique that has proven most effective at bridging this gap is Reinforcement Learning from Human Feedback (RLHF), and at its heart lies a powerful algorithm: Proximal Policy Optimization (PPO).
|
||||
You may have seen diagrams like the one below, which outlines the RLHF training process. It can look intimidating, with a web of interconnected models, losses, and data flows."><meta property="og:url" content="/posts/a-deep-dive-into-ppo-for-language-models/"><meta property="og:site_name" content="Eric X. Liu's Personal Page"><meta property="og:title" content="A Deep Dive into PPO for Language Models"><meta property="og:description" content="Large Language Models (LLMs) have demonstrated astonishing capabilities, but out-of-the-box, they are simply powerful text predictors. They don’t inherently understand what makes a response helpful, harmless, or aligned with human values. The technique that has proven most effective at bridging this gap is Reinforcement Learning from Human Feedback (RLHF), and at its heart lies a powerful algorithm: Proximal Policy Optimization (PPO).
|
||||
You may have seen diagrams like the one below, which outlines the RLHF training process. It can look intimidating, with a web of interconnected models, losses, and data flows."><meta property="og:locale" content="en"><meta property="og:type" content="article"><meta property="article:section" content="posts"><meta property="article:published_time" content="2025-08-02T00:00:00+00:00"><meta property="article:modified_time" content="2025-08-03T03:28:39+00:00"><link rel=canonical href=/posts/a-deep-dive-into-ppo-for-language-models/><link rel=preload href=/fonts/fa-brands-400.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-regular-400.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-solid-900.woff2 as=font type=font/woff2 crossorigin><link rel=stylesheet href=/css/coder.min.60f552a2c0452fcc0254c54f21c3e0728460c1ae85f97a9c35833a222ef8b884.css integrity="sha256-YPVSosBFL8wCVMVPIcPgcoRgwa6F+XqcNYM6Ii74uIQ=" crossorigin=anonymous media=screen><link rel=stylesheet href=/css/coder-dark.min.a00e6364bacbc8266ad1cc81230774a1397198f8cfb7bcba29b7d6fcb54ce57f.css integrity="sha256-oA5jZLrLyCZq0cyBIwd0oTlxmPjPt7y6KbfW/LVM5X8=" crossorigin=anonymous media=screen><link rel=icon type=image/svg+xml href=/images/favicon.svg sizes=any><link rel=icon type=image/png href=/images/favicon-32x32.png sizes=32x32><link rel=icon type=image/png href=/images/favicon-16x16.png sizes=16x16><link rel=apple-touch-icon href=/images/apple-touch-icon.png><link rel=apple-touch-icon sizes=180x180 href=/images/apple-touch-icon.png><link rel=manifest href=/site.webmanifest><link rel=mask-icon href=/images/safari-pinned-tab.svg color=#5bbad5></head><body class="preload-transitions colorscheme-auto"><div class=float-container><a id=dark-mode-toggle class=colorscheme-toggle><i class="fa-solid fa-adjust fa-fw" aria-hidden=true></i></a></div><main class=wrapper><nav class=navigation><section class=container><a class=navigation-title href=/>Eric X. Liu's Personal Page
|
||||
You may have seen diagrams like the one below, which outlines the RLHF training process. It can look intimidating, with a web of interconnected models, losses, and data flows."><meta property="og:locale" content="en"><meta property="og:type" content="article"><meta property="article:section" content="posts"><meta property="article:published_time" content="2025-08-02T00:00:00+00:00"><meta property="article:modified_time" content="2025-08-03T03:28:39+00:00"><link rel=canonical href=/posts/a-deep-dive-into-ppo-for-language-models/><link rel=preload href=/fonts/fa-brands-400.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-regular-400.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-solid-900.woff2 as=font type=font/woff2 crossorigin><link rel=stylesheet href=/css/coder.min.6445a802b9389c9660e1b07b724dcf5718b1065ed2d71b4eeaf981cc7cc5fc46.css integrity="sha256-ZEWoArk4nJZg4bB7ck3PVxixBl7S1xtO6vmBzHzF/EY=" crossorigin=anonymous media=screen><link rel=stylesheet href=/css/coder-dark.min.a00e6364bacbc8266ad1cc81230774a1397198f8cfb7bcba29b7d6fcb54ce57f.css integrity="sha256-oA5jZLrLyCZq0cyBIwd0oTlxmPjPt7y6KbfW/LVM5X8=" crossorigin=anonymous media=screen><link rel=icon type=image/svg+xml href=/images/favicon.svg sizes=any><link rel=icon type=image/png href=/images/favicon-32x32.png sizes=32x32><link rel=icon type=image/png href=/images/favicon-16x16.png sizes=16x16><link rel=apple-touch-icon href=/images/apple-touch-icon.png><link rel=apple-touch-icon sizes=180x180 href=/images/apple-touch-icon.png><link rel=manifest href=/site.webmanifest><link rel=mask-icon href=/images/safari-pinned-tab.svg color=#5bbad5></head><body class="preload-transitions colorscheme-auto"><div class=float-container><a id=dark-mode-toggle class=colorscheme-toggle><i class="fa-solid fa-adjust fa-fw" aria-hidden=true></i></a></div><main class=wrapper><nav class=navigation><section class=container><a class=navigation-title href=/>Eric X. Liu's Personal Page
|
||||
</a><input type=checkbox id=menu-toggle>
|
||||
<label class="menu-button float-right" for=menu-toggle><i class="fa-solid fa-bars fa-fw" aria-hidden=true></i></label><ul class=navigation-list><li class=navigation-item><a class=navigation-link href=/posts/>Posts</a></li><li class=navigation-item><a class=navigation-link href=https://chat.ericxliu.me>Chat</a></li><li class=navigation-item><a class=navigation-link href=https://git.ericxliu.me/user/oauth2/Authenitk>Git</a></li><li class=navigation-item><a class=navigation-link href=https://coder.ericxliu.me/api/v2/users/oidc/callback>Coder</a></li><li class=navigation-item><a class=navigation-link href=https://rss.ericxliu.me/oauth2/oidc/redirect>RSS</a></li><li class=navigation-item><a class=navigation-link href=/>|</a></li><li class=navigation-item><a class=navigation-link href=https://sso.ericxliu.me>Sign in</a></li></ul></section></nav><div class=content><section class="container post"><article><header><div class=post-title><h1 class=title><a class=title-link href=/posts/a-deep-dive-into-ppo-for-language-models/>A Deep Dive into PPO for Language Models</a></h1></div><div class=post-meta><div class=date><span class=posted-on><i class="fa-solid fa-calendar" aria-hidden=true></i>
|
||||
<time datetime=2025-08-02T00:00:00Z>August 2, 2025
|
||||
@@ -19,8 +19,8 @@ where <code>δ_t = r_t + γV(s_{t+1}) - V(s_t)</code></p><ul><li><strong>γ (gam
|
||||
<a class=heading-link href=#avoiding-amnesia-the-pretraining-loss><i class="fa-solid fa-link" aria-hidden=true title="Link to heading"></i>
|
||||
<span class=sr-only>Link to heading</span></a></h3><p>There’s one final problem. If we only optimize for the PPO loss, the model might learn to “hack” the reward model by generating repetitive or nonsensical text that gets a high score. In doing so, it could suffer from <strong>catastrophic forgetting</strong>, losing its fundamental grasp of grammar and facts.</p><p>To prevent this, we introduce a second loss term. As seen in the diagram, we mix in data from the original <strong>Pretraining Data</strong> (or the dataset used for Supervised Fine-Tuning). We calculate a standard next-token prediction loss (<code>LM Loss</code>) on this high-quality data.</p><p>The final loss for the Actor is a combination of both objectives:</p><p><strong>Total Loss = Loss_PPO + <code>λ_ptx</code> * Loss_LM</strong></p><p>This brilliantly balances two goals:</p><ol><li>The <code>Loss_PPO</code> pushes the model towards behaviors that align with human preferences.</li><li>The <code>Loss_LM</code> acts as a regularizer, pulling the model back towards its core language capabilities and preventing it from drifting into gibberish.</li></ol><h3 id=the-full-training-loop>The Full Training Loop
|
||||
<a class=heading-link href=#the-full-training-loop><i class="fa-solid fa-link" aria-hidden=true title="Link to heading"></i>
|
||||
<span class=sr-only>Link to heading</span></a></h3><p>Now, we can assemble the entire process into a clear, iterative loop:</p><ol><li><strong>Collect</strong>: The current Actor policy <code>π_k</code> generates responses to a batch of prompts. These experiences—<code>(state, action, probability, reward, value)</code>—are stored in an <strong>Experience Buffer</strong>.</li><li><strong>Calculate</strong>: Once the buffer is full, we use the collected data to compute the advantage estimates <code>Â_t</code> for every single token-generation step.</li><li><strong>Optimize</strong>: For a few epochs, we repeatedly sample mini-batches from the buffer and update the Actor and Critic models. The Actor is updated using the combined <code>PPO-clip Loss</code> and <code>LM Loss</code>. The Critic is updated to improve its value predictions.</li><li><strong>Flush and Repeat</strong>: After the optimization phase, the entire experience buffer is discarded. The data is now “stale” because our policy has changed. The newly updated policy <code>π_{k+1}</code> becomes the new Actor, and we return to step 1 to collect fresh data.</li></ol><p>This cycle of collection and optimization allows the language model to gradually and safely steer its behavior towards human-defined goals, creating the helpful and aligned AI assistants we interact with today.</p><hr><p><strong>References:</strong></p><ol><li>Schulman, J., Wolski, F., Dhariwal, P., Radford, A., & Klimov, O. (2017). <em>Proximal Policy Optimization Algorithms</em>. arXiv preprint arXiv:1707.06347.</li><li>Schulman, J., Moritz, P., Levine, S., Jordan, M., & Abbeel, P. (2015). <em>High-Dimensional Continuous Control Using Generalized Advantage Estimation</em>. arXiv preprint arXiv:1506.02438.</li><li>Ouyang, L., et al. (2022). <em>Training language models to follow instructions with human feedback</em>. Advances in Neural Information Processing Systems 35.</li></ol></div><footer></footer></article></section></div><footer class=footer><section class=container>©
|
||||
<span class=sr-only>Link to heading</span></a></h3><p>Now, we can assemble the entire process into a clear, iterative loop:</p><ol><li><strong>Collect</strong>: The current Actor policy <code>π_k</code> generates responses to a batch of prompts. These experiences—<code>(state, action, probability, reward, value)</code>—are stored in an <strong>Experience Buffer</strong>.</li><li><strong>Calculate</strong>: Once the buffer is full, we use the collected data to compute the advantage estimates <code>Â_t</code> for every single token-generation step.</li><li><strong>Optimize</strong>: For a few epochs, we repeatedly sample mini-batches from the buffer and update the Actor and Critic models. The Actor is updated using the combined <code>PPO-clip Loss</code> and <code>LM Loss</code>. The Critic is updated to improve its value predictions.</li><li><strong>Flush and Repeat</strong>: After the optimization phase, the entire experience buffer is discarded. The data is now “stale” because our policy has changed. The newly updated policy <code>π_{k+1}</code> becomes the new Actor, and we return to step 1 to collect fresh data.</li></ol><p>This cycle of collection and optimization allows the language model to gradually and safely steer its behavior towards human-defined goals, creating the helpful and aligned AI assistants we interact with today.</p><hr><p><strong>References:</strong></p><ol><li>Schulman, J., Wolski, F., Dhariwal, P., Radford, A., & Klimov, O. (2017). <em>Proximal Policy Optimization Algorithms</em>. arXiv preprint arXiv:1707.06347.</li><li>Schulman, J., Moritz, P., Levine, S., Jordan, M., & Abbeel, P. (2015). <em>High-Dimensional Continuous Control Using Generalized Advantage Estimation</em>. arXiv preprint arXiv:1506.02438.</li><li>Ouyang, L., et al. (2022). <em>Training language models to follow instructions with human feedback</em>. Advances in Neural Information Processing Systems 35.</li></ol></div><footer><div id=disqus_thread></div><script>window.disqus_config=function(){},function(){if(["localhost","127.0.0.1"].indexOf(window.location.hostname)!=-1){document.getElementById("disqus_thread").innerHTML="Disqus comments not available by default when the website is previewed locally.";return}var t=document,e=t.createElement("script");e.async=!0,e.src="//ericxliu-me.disqus.com/embed.js",e.setAttribute("data-timestamp",+new Date),(t.head||t.body).appendChild(e)}(),document.addEventListener("themeChanged",function(){document.readyState=="complete"&&DISQUS.reset({reload:!0,config:disqus_config})})</script></footer></article></section></div><footer class=footer><section class=container>©
|
||||
2016 -
|
||||
2025
|
||||
Eric X. Liu
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/fd19c59">[fd19c59]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script></body></html>
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/b4cc012">[b4cc012]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script></body></html>
|
||||
Reference in New Issue
Block a user