Buckets:
| <meta charset="utf-8" /><meta name="hf:doc:metadata" content="{"title":"GRPO Trainer","local":"grpo-trainer","sections":[{"title":"Overview","local":"overview","sections":[],"depth":2},{"title":"Quick start","local":"quick-start","sections":[],"depth":2},{"title":"Looking deeper into the GRPO method","local":"looking-deeper-into-the-grpo-method","sections":[{"title":"Generating completions","local":"generating-completions","sections":[],"depth":3},{"title":"Computing the advantage","local":"computing-the-advantage","sections":[],"depth":3},{"title":"Estimating the KL divergence","local":"estimating-the-kl-divergence","sections":[],"depth":3},{"title":"Computing the loss","local":"computing-the-loss","sections":[{"title":"Loss Types","local":"loss-types","sections":[],"depth":4}],"depth":3}],"depth":2},{"title":"Logged metrics","local":"logged-metrics","sections":[],"depth":2},{"title":"Customization","local":"customization","sections":[{"title":"Speed up training with vLLM-powered generation","local":"speed-up-training-with-vllm-powered-generation","sections":[{"title":"🔌 Option 1: Server mode","local":"-option-1-server-mode","sections":[],"depth":4},{"title":"🧩 Option 2: Colocate mode","local":"-option-2-colocate-mode","sections":[],"depth":4}],"depth":3},{"title":"GRPO at scale: train a 70B+ Model on multiple nodes","local":"grpo-at-scale-train-a-70b-model-on-multiple-nodes","sections":[],"depth":3},{"title":"Using a custom reward function","local":"using-a-custom-reward-function","sections":[{"title":"Example 1: Reward longer completions","local":"example-1-reward-longer-completions","sections":[],"depth":4},{"title":"Example 1.1: Reward longer completions (based on the number of characters)","local":"example-11-reward-longer-completions-based-on-the-number-of-characters","sections":[],"depth":4},{"title":"Example 2: Reward completions with a specific format","local":"example-2-reward-completions-with-a-specific-format","sections":[],"depth":4},{"title":"Example 3: Reward completions based on a reference","local":"example-3-reward-completions-based-on-a-reference","sections":[],"depth":4},{"title":"Example 4: Multi-task reward functions","local":"example-4-multi-task-reward-functions","sections":[],"depth":4},{"title":"Passing the reward function to the trainer","local":"passing-the-reward-function-to-the-trainer","sections":[],"depth":4}],"depth":3}],"depth":2},{"title":"Vision-Language Model (VLM) Training","local":"vision-language-model-vlm-training","sections":[{"title":"Supported Models","local":"supported-models","sections":[],"depth":3},{"title":"Quick Start","local":"quick-start","sections":[],"depth":3},{"title":"Configuration Tips","local":"configuration-tips","sections":[],"depth":3},{"title":"Dataset Format","local":"dataset-format","sections":[],"depth":3}],"depth":2},{"title":"GRPOTrainer","local":"trl.GRPOTrainer","sections":[],"depth":2},{"title":"GRPOConfig","local":"trl.GRPOConfig","sections":[],"depth":2}],"depth":1}"> | |
| <link href="/docs/trl/pr_4331/en/_app/immutable/assets/0.e3b0c442.css" rel="modulepreload"> | |
| <link rel="modulepreload" href="/docs/trl/pr_4331/en/_app/immutable/entry/start.6bbbc54b.js"> | |
| <link rel="modulepreload" href="/docs/trl/pr_4331/en/_app/immutable/chunks/scheduler.7b731bd4.js"> | |
| <link rel="modulepreload" href="/docs/trl/pr_4331/en/_app/immutable/chunks/singletons.55eb59f9.js"> | |
| <link rel="modulepreload" href="/docs/trl/pr_4331/en/_app/immutable/chunks/index.ac28c20f.js"> | |
| <link rel="modulepreload" href="/docs/trl/pr_4331/en/_app/immutable/chunks/paths.677b038d.js"> | |
| <link rel="modulepreload" href="/docs/trl/pr_4331/en/_app/immutable/entry/app.b003256e.js"> | |
| <link rel="modulepreload" href="/docs/trl/pr_4331/en/_app/immutable/chunks/preload-helper.71df5523.js"> | |
| <link rel="modulepreload" href="/docs/trl/pr_4331/en/_app/immutable/chunks/index.cc268345.js"> | |
| <link rel="modulepreload" href="/docs/trl/pr_4331/en/_app/immutable/nodes/0.c996cd3a.js"> | |
| <link rel="modulepreload" href="/docs/trl/pr_4331/en/_app/immutable/chunks/each.e59479a4.js"> | |
| <link rel="modulepreload" href="/docs/trl/pr_4331/en/_app/immutable/nodes/19.08bc5f45.js"> | |
| <link rel="modulepreload" href="/docs/trl/pr_4331/en/_app/immutable/chunks/MermaidChart.svelte_svelte_type_style_lang.d403d039.js"> | |
| <link rel="modulepreload" href="/docs/trl/pr_4331/en/_app/immutable/chunks/Docstring.59783212.js"> | |
| <link rel="modulepreload" href="/docs/trl/pr_4331/en/_app/immutable/chunks/CodeBlock.17bc4142.js"> | |
| <link rel="modulepreload" href="/docs/trl/pr_4331/en/_app/immutable/chunks/ExampleCodeBlock.b6fe85ba.js"><!-- HEAD_svelte-u9bgzb_START --><meta name="hf:doc:metadata" content="{"title":"GRPO Trainer","local":"grpo-trainer","sections":[{"title":"Overview","local":"overview","sections":[],"depth":2},{"title":"Quick start","local":"quick-start","sections":[],"depth":2},{"title":"Looking deeper into the GRPO method","local":"looking-deeper-into-the-grpo-method","sections":[{"title":"Generating completions","local":"generating-completions","sections":[],"depth":3},{"title":"Computing the advantage","local":"computing-the-advantage","sections":[],"depth":3},{"title":"Estimating the KL divergence","local":"estimating-the-kl-divergence","sections":[],"depth":3},{"title":"Computing the loss","local":"computing-the-loss","sections":[{"title":"Loss Types","local":"loss-types","sections":[],"depth":4}],"depth":3}],"depth":2},{"title":"Logged metrics","local":"logged-metrics","sections":[],"depth":2},{"title":"Customization","local":"customization","sections":[{"title":"Speed up training with vLLM-powered generation","local":"speed-up-training-with-vllm-powered-generation","sections":[{"title":"🔌 Option 1: Server mode","local":"-option-1-server-mode","sections":[],"depth":4},{"title":"🧩 Option 2: Colocate mode","local":"-option-2-colocate-mode","sections":[],"depth":4}],"depth":3},{"title":"GRPO at scale: train a 70B+ Model on multiple nodes","local":"grpo-at-scale-train-a-70b-model-on-multiple-nodes","sections":[],"depth":3},{"title":"Using a custom reward function","local":"using-a-custom-reward-function","sections":[{"title":"Example 1: Reward longer completions","local":"example-1-reward-longer-completions","sections":[],"depth":4},{"title":"Example 1.1: Reward longer completions (based on the number of characters)","local":"example-11-reward-longer-completions-based-on-the-number-of-characters","sections":[],"depth":4},{"title":"Example 2: Reward completions with a specific format","local":"example-2-reward-completions-with-a-specific-format","sections":[],"depth":4},{"title":"Example 3: Reward completions based on a reference","local":"example-3-reward-completions-based-on-a-reference","sections":[],"depth":4},{"title":"Example 4: Multi-task reward functions","local":"example-4-multi-task-reward-functions","sections":[],"depth":4},{"title":"Passing the reward function to the trainer","local":"passing-the-reward-function-to-the-trainer","sections":[],"depth":4}],"depth":3}],"depth":2},{"title":"Vision-Language Model (VLM) Training","local":"vision-language-model-vlm-training","sections":[{"title":"Supported Models","local":"supported-models","sections":[],"depth":3},{"title":"Quick Start","local":"quick-start","sections":[],"depth":3},{"title":"Configuration Tips","local":"configuration-tips","sections":[],"depth":3},{"title":"Dataset Format","local":"dataset-format","sections":[],"depth":3}],"depth":2},{"title":"GRPOTrainer","local":"trl.GRPOTrainer","sections":[],"depth":2},{"title":"GRPOConfig","local":"trl.GRPOConfig","sections":[],"depth":2}],"depth":1}"><!-- HEAD_svelte-u9bgzb_END --> <p></p> <div class="items-center shrink-0 min-w-[100px] max-sm:min-w-[50px] justify-end ml-auto flex" style="float: right; margin-left: 10px; display: inline-flex; position: relative; z-index: 10;"><div class="inline-flex rounded-md max-sm:rounded-sm"><button class="inline-flex items-center gap-1 max-sm:gap-0.5 h-6 max-sm:h-5 px-2 max-sm:px-1.5 text-[11px] max-sm:text-[9px] font-medium text-gray-800 border border-r-0 rounded-l-md max-sm:rounded-l-sm border-gray-200 bg-white hover:shadow-inner dark:border-gray-850 dark:bg-gray-950 dark:text-gray-200 dark:hover:bg-gray-800" aria-live="polite"><span class="inline-flex items-center justify-center rounded-md p-0.5 max-sm:p-0"><svg class="w-3 h-3 max-sm:w-2.5 max-sm:h-2.5" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg></span> <span>Copy page</span></button> <button class="inline-flex items-center justify-center w-6 max-sm:w-5 h-6 max-sm:h-5 disabled:pointer-events-none text-sm text-gray-500 hover:text-gray-700 dark:hover:text-white rounded-r-md max-sm:rounded-r-sm border border-l transition border-gray-200 bg-white hover:shadow-inner dark:border-gray-850 dark:bg-gray-950 dark:text-gray-200 dark:hover:bg-gray-800" aria-haspopup="menu" aria-expanded="false" aria-label="Open copy menu"><svg class="transition-transform text-gray-400 overflow-visible w-3 h-3 max-sm:w-2.5 max-sm:h-2.5 rotate-0" width="1em" height="1em" viewBox="0 0 12 7" fill="none" xmlns="http://www.w3.org/2000/svg"><path d="M1 1L6 6L11 1" stroke="currentColor"></path></svg></button></div> </div> <h1 class="relative group"><a id="grpo-trainer" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#grpo-trainer"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>GRPO Trainer</span></h1> <p data-svelte-h="svelte-5f4e9w"><a href="https://huggingface.co/models?other=grpo,trl" rel="nofollow"><img src="https://img.shields.io/badge/All_models-GRPO-blue" alt="model badge"></a></p> <h2 class="relative group"><a id="overview" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#overview"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Overview</span></h2> <p data-svelte-h="svelte-1btigzj">TRL supports the GRPO Trainer for training language models, as described in the paper <a href="https://huggingface.co/papers/2402.03300" rel="nofollow">DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models</a> by <a href="https://huggingface.co/syhia" rel="nofollow">Zhihong Shao</a>, <a href="https://huggingface.co/peiyiwang89" rel="nofollow">Peiyi Wang</a>, <a href="https://huggingface.co/zqh11" rel="nofollow">Qihao Zhu</a>, Runxin Xu, <a href="https://huggingface.co/haha-point" rel="nofollow">Junxiao Song</a>, Mingchuan Zhang, Y. K. Li, Y. Wu, <a href="https://huggingface.co/guoday" rel="nofollow">Daya Guo</a>.</p> <p data-svelte-h="svelte-vfdo9a">The abstract from the paper is the following:</p> <blockquote data-svelte-h="svelte-maid5i"><p>Mathematical reasoning poses a significant challenge for language models due to its complex and structured nature. In this paper, we introduce DeepSeekMath 7B, which continues pre-training DeepSeek-Coder-Base-v1.5 7B with 120B math-related tokens sourced from Common Crawl, together with natural language and code data. DeepSeekMath 7B has achieved an impressive score of 51.7% on the competition-level MATH benchmark without relying on external toolkits and voting techniques, approaching the performance level of Gemini-Ultra and GPT-4. Self-consistency over 64 samples from DeepSeekMath 7B achieves 60.9% on MATH. The mathematical reasoning capability of DeepSeekMath is attributed to two key factors: First, we harness the significant potential of publicly available web data through a meticulously engineered data selection pipeline. Second, we introduce Group Relative Policy Optimization (GRPO), a variant of Proximal Policy Optimization (PPO), that enhances mathematical reasoning abilities while concurrently optimizing the memory usage of PPO.</p></blockquote> <p data-svelte-h="svelte-mxzj2f">This post-training method was contributed by <a href="https://huggingface.co/qgallouedec" rel="nofollow">Quentin Gallouédec</a>.</p> <h2 class="relative group"><a id="quick-start" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#quick-start"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Quick start</span></h2> <p data-svelte-h="svelte-lovg5j">This example demonstrates how to train a model using the GRPO method. We train a <a href="https://huggingface.co/Qwen/Qwen2-0.5B-Instruct" rel="nofollow">Qwen 0.5B Instruct model</a> with the prompts from the <a href="https://huggingface.co/datasets/trl-lib/ultrafeedback-prompt" rel="nofollow">UltraFeedback prompts dataset</a>. You can view the data in the dataset here:</p> <iframe src="https://huggingface.co/datasets/trl-lib/ultrafeedback-prompt/embed/viewer/default/train?row=0" frameborder="0" width="100%" height="560px"></iframe> <p data-svelte-h="svelte-fsw01e">Below is the script to train the model.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-comment"># train_grpo.py</span> | |
| <span class="hljs-keyword">from</span> datasets <span class="hljs-keyword">import</span> load_dataset | |
| <span class="hljs-keyword">from</span> trl <span class="hljs-keyword">import</span> GRPOConfig, GRPOTrainer | |
| dataset = load_dataset(<span class="hljs-string">"trl-lib/ultrafeedback-prompt"</span>, split=<span class="hljs-string">"train"</span>) | |
| <span class="hljs-comment"># Dummy reward function for demonstration purposes</span> | |
| <span class="hljs-keyword">def</span> <span class="hljs-title function_">reward_num_unique_letters</span>(<span class="hljs-params">completions, **kwargs</span>): | |
| <span class="hljs-string">"""Reward function that rewards completions with more unique letters."""</span> | |
| completion_contents = [completion[<span class="hljs-number">0</span>][<span class="hljs-string">"content"</span>] <span class="hljs-keyword">for</span> completion <span class="hljs-keyword">in</span> completions] | |
| <span class="hljs-keyword">return</span> [<span class="hljs-built_in">float</span>(<span class="hljs-built_in">len</span>(<span class="hljs-built_in">set</span>(content))) <span class="hljs-keyword">for</span> content <span class="hljs-keyword">in</span> completion_contents] | |
| training_args = GRPOConfig(output_dir=<span class="hljs-string">"Qwen2-0.5B-GRPO"</span>) | |
| trainer = GRPOTrainer( | |
| model=<span class="hljs-string">"Qwen/Qwen2-0.5B-Instruct"</span>, | |
| reward_funcs=reward_num_unique_letters, | |
| args=training_args, | |
| train_dataset=dataset, | |
| ) | |
| trainer.train()<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-15hino8">Execute the script using the following command:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->accelerate launch train_grpo.py<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1p32u9o">Distributed across 8 GPUs, the training takes approximately 1 day.</p> <p data-svelte-h="svelte-19pqwae"><img src="https://huggingface.co/datasets/trl-lib/documentation-images/resolve/main/grpo_curves.png" alt="GRPO curves"></p> <h2 class="relative group"><a id="looking-deeper-into-the-grpo-method" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#looking-deeper-into-the-grpo-method"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Looking deeper into the GRPO method</span></h2> <p data-svelte-h="svelte-ir94jd">GRPO is an online learning algorithm, meaning it improves iteratively by using the data generated by the trained model itself during training. The intuition behind GRPO objective is to maximize the advantage of the generated completions, while ensuring that the model remains close to the reference policy. To understand how GRPO works, it can be broken down into four main steps: <strong>Generating completions</strong>, <strong>computing the advantage</strong>, <strong>estimating the KL divergence</strong>, and <strong>computing the loss</strong>.</p> <p data-svelte-h="svelte-sxakza"><img src="https://huggingface.co/datasets/trl-lib/documentation-images/resolve/main/grpo_visual.png" alt="GRPO visual"></p> <h3 class="relative group"><a id="generating-completions" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#generating-completions"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Generating completions</span></h3> <p>At each training step, we sample a batch of prompts and generate a set of <!-- HTML_TAG_START --><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>G</mi></mrow><annotation encoding="application/x-tex"> G </annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.6833em;"></span><span class="mord mathnormal">G</span></span></span></span><!-- HTML_TAG_END --> completions for each prompt (denoted as <!-- HTML_TAG_START --><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mi>o</mi><mi>i</mi></msub></mrow><annotation encoding="application/x-tex"> o_i </annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.5806em;vertical-align:-0.15em;"></span><span class="mord"><span class="mord mathnormal">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight">i</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span></span></span></span><!-- HTML_TAG_END -->).</p> <h3 class="relative group"><a id="computing-the-advantage" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#computing-the-advantage"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Computing the advantage</span></h3> <p>For each of the <!-- HTML_TAG_START --><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>G</mi></mrow><annotation encoding="application/x-tex"> G </annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.6833em;"></span><span class="mord mathnormal">G</span></span></span></span><!-- HTML_TAG_END --> sequences, we compute the reward using a reward model or reward function. To align with the comparative nature of reward models—typically trained on datasets of comparisons between outputs for the same question—the advantage is calculated to reflect these relative comparisons. It is normalized as follows: | |
| <!-- HTML_TAG_START --><span class="katex-display"><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML" display="block"><semantics><mrow><msub><mover accent="true"><mi>A</mi><mo>^</mo></mover><mrow><mi>i</mi><mo separator="true">,</mo><mi>t</mi></mrow></msub><mo>=</mo><mfrac><mrow><msub><mi>r</mi><mi>i</mi></msub><mo>−</mo><mtext>mean</mtext><mo stretchy="false">(</mo><mi mathvariant="bold">r</mi><mo stretchy="false">)</mo></mrow><mrow><mtext>std</mtext><mo stretchy="false">(</mo><mi mathvariant="bold">r</mi><mo stretchy="false">)</mo></mrow></mfrac></mrow><annotation encoding="application/x-tex">\hat{A}_{i,t} = \frac{r_i - \text{mean}(\mathbf{r})}{\text{std}(\mathbf{r})}</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:1.2329em;vertical-align:-0.2861em;"></span><span class="mord"><span class="mord accent"><span class="vlist-t"><span class="vlist-r"><span class="vlist" style="height:0.9468em;"><span style="top:-3em;"><span class="pstrut" style="height:3em;"></span><span class="mord mathnormal">A</span></span><span style="top:-3.2523em;"><span class="pstrut" style="height:3em;"></span><span class="accent-body" style="left:-0.1111em;"><span class="mord">^</span></span></span></span></span></span></span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:2.363em;vertical-align:-0.936em;"></span><span class="mord"><span class="mopen nulldelimiter"></span><span class="mfrac"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.427em;"><span style="top:-2.314em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord text"><span class="mord">std</span></span><span class="mopen">(</span><span class="mord mathbf">r</span><span class="mclose">)</span></span></span><span style="top:-3.23em;"><span class="pstrut" style="height:3em;"></span><span class="frac-line" style="border-bottom-width:0.04em;"></span></span><span style="top:-3.677em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord"><span class="mord mathnormal" style="margin-right:0.02778em;">r</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:-0.0278em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight">i</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mbin">−</span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mord text"><span class="mord">mean</span></span><span class="mopen">(</span><span class="mord mathbf">r</span><span class="mclose">)</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.936em;"><span></span></span></span></span></span><span class="mclose nulldelimiter"></span></span></span></span></span></span><!-- HTML_TAG_END --></p> <p data-svelte-h="svelte-1gfu28r">This approach gives the method its name: <strong>Group Relative Policy Optimization (GRPO)</strong>.</p> <blockquote class="tip"><p>It was shown in the paper <a href="https://huggingface.co/papers/2503.20783" rel="nofollow" data-svelte-h="svelte-u44a8w">Understanding R1-Zero-Like Training: A Critical Perspective</a> that scaling by <!-- HTML_TAG_START --><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mtext>std</mtext><mo stretchy="false">(</mo><mi mathvariant="bold">r</mi><mo stretchy="false">)</mo></mrow><annotation encoding="application/x-tex"> \text{std}(\mathbf{r}) </annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord text"><span class="mord">std</span></span><span class="mopen">(</span><span class="mord mathbf">r</span><span class="mclose">)</span></span></span></span><!-- HTML_TAG_END --> may cause a question-level difficulty bias. You can disable this scaling by setting <code data-svelte-h="svelte-bmfm87">scale_rewards=False</code> in <a href="/docs/trl/pr_4331/en/grpo_trainer#trl.GRPOConfig" data-svelte-h="svelte-jirtk">GRPOConfig</a>.</p></blockquote> <blockquote class="tip" data-svelte-h="svelte-92i8ao"><p>As shown in <a href="https://huggingface.co/papers/2508.08221" rel="nofollow">Part I: Tricks or Traps? A Deep Dive into RL for LLM Reasoning (Lite PPO)</a>, calculating the mean at the local (group) level and the standard deviation at the global (batch) level enables more robust reward shaping. You can use this scaling strategy by setting <code>scale_rewards="batch"</code> in <a href="/docs/trl/pr_4331/en/grpo_trainer#trl.GRPOConfig">GRPOConfig</a>.</p></blockquote> <h3 class="relative group"><a id="estimating-the-kl-divergence" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#estimating-the-kl-divergence"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Estimating the KL divergence</span></h3> <p>KL divergence is estimated using the approximator introduced by <a href="http://joschu.net/blog/kl-approx.html" rel="nofollow" data-svelte-h="svelte-77j2bq">Schulman et al. (2020)</a>. The approximator is defined as follows: | |
| <!-- HTML_TAG_START --><span class="katex-display"><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML" display="block"><semantics><mrow><msub><mi mathvariant="double-struck">D</mi><mtext>KL</mtext></msub><mrow><mo fence="true">[</mo><msub><mi>π</mi><mi>θ</mi></msub><mi mathvariant="normal">∥</mi><msub><mi>π</mi><mtext>ref</mtext></msub><mo fence="true">]</mo></mrow><mo>=</mo><mfrac><mrow><msub><mi>π</mi><mtext>ref</mtext></msub><mo stretchy="false">(</mo><msub><mi>o</mi><mrow><mi>i</mi><mo separator="true">,</mo><mi>t</mi></mrow></msub><mo>∣</mo><mi>q</mi><mo separator="true">,</mo><msub><mi>o</mi><mrow><mi>i</mi><mo separator="true">,</mo><mo><</mo><mi>t</mi></mrow></msub><mo stretchy="false">)</mo></mrow><mrow><msub><mi>π</mi><mi>θ</mi></msub><mo stretchy="false">(</mo><msub><mi>o</mi><mrow><mi>i</mi><mo separator="true">,</mo><mi>t</mi></mrow></msub><mo>∣</mo><mi>q</mi><mo separator="true">,</mo><msub><mi>o</mi><mrow><mi>i</mi><mo separator="true">,</mo><mo><</mo><mi>t</mi></mrow></msub><mo stretchy="false">)</mo></mrow></mfrac><mo>−</mo><mi>log</mi><mo></mo><mfrac><mrow><msub><mi>π</mi><mtext>ref</mtext></msub><mo stretchy="false">(</mo><msub><mi>o</mi><mrow><mi>i</mi><mo separator="true">,</mo><mi>t</mi></mrow></msub><mo>∣</mo><mi>q</mi><mo separator="true">,</mo><msub><mi>o</mi><mrow><mi>i</mi><mo separator="true">,</mo><mo><</mo><mi>t</mi></mrow></msub><mo stretchy="false">)</mo></mrow><mrow><msub><mi>π</mi><mi>θ</mi></msub><mo stretchy="false">(</mo><msub><mi>o</mi><mrow><mi>i</mi><mo separator="true">,</mo><mi>t</mi></mrow></msub><mo>∣</mo><mi>q</mi><mo separator="true">,</mo><msub><mi>o</mi><mrow><mi>i</mi><mo separator="true">,</mo><mo><</mo><mi>t</mi></mrow></msub><mo stretchy="false">)</mo></mrow></mfrac><mo>−</mo><mn>1</mn><mo separator="true">,</mo></mrow><annotation encoding="application/x-tex">\mathbb{D}_{\text{KL}}\left[\pi_\theta \|\pi_{\text{ref}}\right] = \frac{\pi_{\text{ref}}(o_{i,t} \mid q, o_{i,<t})}{\pi_\theta(o_{i,t} \mid q, o_{i,<t})} - \log \frac{\pi_{\text{ref}}(o_{i,t} \mid q, o_{i,<t})}{\pi_\theta(o_{i,t} \mid q, o_{i,<t})} - 1, | |
| </annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord"><span class="mord mathbb">D</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3283em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord text mtight"><span class="mord mtight">KL</span></span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.1667em;"></span><span class="minner"><span class="mopen delimcenter" style="top:0em;">[</span><span class="mord"><span class="mord mathnormal" style="margin-right:0.03588em;">π</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3361em;"><span style="top:-2.55em;margin-left:-0.0359em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight" style="margin-right:0.02778em;">θ</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mord">∥</span><span class="mord"><span class="mord mathnormal" style="margin-right:0.03588em;">π</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3361em;"><span style="top:-2.55em;margin-left:-0.0359em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord text mtight"><span class="mord mtight">ref</span></span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mclose delimcenter" style="top:0em;">]</span></span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:2.3991em;vertical-align:-0.9721em;"></span><span class="mord"><span class="mopen nulldelimiter"></span><span class="mfrac"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.427em;"><span style="top:-2.314em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord"><span class="mord mathnormal" style="margin-right:0.03588em;">π</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3361em;"><span style="top:-2.55em;margin-left:-0.0359em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight" style="margin-right:0.02778em;">θ</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mopen">(</span><span class="mord"><span class="mord mathnormal">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">∣</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mord mathnormal" style="margin-right:0.03588em;">q</span><span class="mpunct">,</span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mord"><span class="mord mathnormal">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mrel mtight"><</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mclose">)</span></span></span><span style="top:-3.23em;"><span class="pstrut" style="height:3em;"></span><span class="frac-line" style="border-bottom-width:0.04em;"></span></span><span style="top:-3.677em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord"><span class="mord mathnormal" style="margin-right:0.03588em;">π</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3361em;"><span style="top:-2.55em;margin-left:-0.0359em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord text mtight"><span class="mord mtight">ref</span></span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mopen">(</span><span class="mord"><span class="mord mathnormal">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">∣</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mord mathnormal" style="margin-right:0.03588em;">q</span><span class="mpunct">,</span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mord"><span class="mord mathnormal">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mrel mtight"><</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mclose">)</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.9721em;"><span></span></span></span></span></span><span class="mclose nulldelimiter"></span></span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mbin">−</span><span class="mspace" style="margin-right:0.2222em;"></span></span><span class="base"><span class="strut" style="height:2.3991em;vertical-align:-0.9721em;"></span><span class="mop">lo<span style="margin-right:0.01389em;">g</span></span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mord"><span class="mopen nulldelimiter"></span><span class="mfrac"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.427em;"><span style="top:-2.314em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord"><span class="mord mathnormal" style="margin-right:0.03588em;">π</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3361em;"><span style="top:-2.55em;margin-left:-0.0359em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight" style="margin-right:0.02778em;">θ</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mopen">(</span><span class="mord"><span class="mord mathnormal">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">∣</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mord mathnormal" style="margin-right:0.03588em;">q</span><span class="mpunct">,</span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mord"><span class="mord mathnormal">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mrel mtight"><</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mclose">)</span></span></span><span style="top:-3.23em;"><span class="pstrut" style="height:3em;"></span><span class="frac-line" style="border-bottom-width:0.04em;"></span></span><span style="top:-3.677em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord"><span class="mord mathnormal" style="margin-right:0.03588em;">π</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3361em;"><span style="top:-2.55em;margin-left:-0.0359em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord text mtight"><span class="mord mtight">ref</span></span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mopen">(</span><span class="mord"><span class="mord mathnormal">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">∣</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mord mathnormal" style="margin-right:0.03588em;">q</span><span class="mpunct">,</span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mord"><span class="mord mathnormal">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mrel mtight"><</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mclose">)</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.9721em;"><span></span></span></span></span></span><span class="mclose nulldelimiter"></span></span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mbin">−</span><span class="mspace" style="margin-right:0.2222em;"></span></span><span class="base"><span class="strut" style="height:0.8389em;vertical-align:-0.1944em;"></span><span class="mord">1</span><span class="mpunct">,</span></span></span></span></span><!-- HTML_TAG_END --></p> <h3 class="relative group"><a id="computing-the-loss" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#computing-the-loss"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Computing the loss</span></h3> <p>The objective is to maximize the advantage while ensuring that the model remains close to the reference policy. Consequently, the loss is defined as follows: | |
| <!-- HTML_TAG_START --><span class="katex-display"><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML" display="block"><semantics><mrow><msub><mi mathvariant="script">L</mi><mtext>GRPO</mtext></msub><mo stretchy="false">(</mo><mi>θ</mi><mo stretchy="false">)</mo><mo>=</mo><mo>−</mo><mfrac><mn>1</mn><mrow><munderover><mo>∑</mo><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><mi>G</mi></munderover><mi mathvariant="normal">∣</mi><msub><mi>o</mi><mi>i</mi></msub><mi mathvariant="normal">∣</mi></mrow></mfrac><munderover><mo>∑</mo><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><mi>G</mi></munderover><munderover><mo>∑</mo><mrow><mi>t</mi><mo>=</mo><mn>1</mn></mrow><mrow><mi mathvariant="normal">∣</mi><msub><mi>o</mi><mi>i</mi></msub><mi mathvariant="normal">∣</mi></mrow></munderover><mrow><mo fence="true">[</mo><mfrac><mrow><msub><mi>π</mi><mi>θ</mi></msub><mo stretchy="false">(</mo><msub><mi>o</mi><mrow><mi>i</mi><mo separator="true">,</mo><mi>t</mi></mrow></msub><mo>∣</mo><mi>q</mi><mo separator="true">,</mo><msub><mi>o</mi><mrow><mi>i</mi><mo separator="true">,</mo><mo><</mo><mi>t</mi></mrow></msub><mo stretchy="false">)</mo></mrow><msub><mrow><mo fence="true">[</mo><msub><mi>π</mi><mi>θ</mi></msub><mo stretchy="false">(</mo><msub><mi>o</mi><mrow><mi>i</mi><mo separator="true">,</mo><mi>t</mi></mrow></msub><mo>∣</mo><mi>q</mi><mo separator="true">,</mo><msub><mi>o</mi><mrow><mi>i</mi><mo separator="true">,</mo><mo><</mo><mi>t</mi></mrow></msub><mo stretchy="false">)</mo><mo fence="true">]</mo></mrow><mtext>no grad</mtext></msub></mfrac><msub><mover accent="true"><mi>A</mi><mo>^</mo></mover><mrow><mi>i</mi><mo separator="true">,</mo><mi>t</mi></mrow></msub><mo>−</mo><mi>β</mi><msub><mi mathvariant="double-struck">D</mi><mtext>KL</mtext></msub><mrow><mo fence="true">[</mo><msub><mi>π</mi><mi>θ</mi></msub><mi mathvariant="normal">∥</mi><msub><mi>π</mi><mtext>ref</mtext></msub><mo fence="true">]</mo></mrow><mo fence="true">]</mo></mrow><mo separator="true">,</mo></mrow><annotation encoding="application/x-tex"> | |
| \mathcal{L}_{\text{GRPO}}(\theta) = -\frac{1}{\sum_{i=1}^G |o_i|} \sum_{i=1}^G \sum_{t=1}^{|o_i|} \left[ \frac{\pi_\theta(o_{i,t} \mid q, o_{i,< t})}{\left[\pi_\theta(o_{i,t} \mid q, o_{i,< t})\right]_{\text{no grad}}} \hat{A}_{i,t} - \beta \mathbb{D}_{\text{KL}}\left[\pi_\theta \| \pi_{\text{ref}}\right] \right], | |
| </annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord"><span class="mord mathcal">L</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3283em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord text mtight"><span class="mord mtight">GRPO</span></span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mopen">(</span><span class="mord mathnormal" style="margin-right:0.02778em;">θ</span><span class="mclose">)</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:3.2387em;vertical-align:-1.2777em;"></span><span class="mord">−</span><span class="mord"><span class="mopen nulldelimiter"></span><span class="mfrac"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.3214em;"><span style="top:-2.1288em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mop"><span class="mop op-symbol small-op" style="position:relative;top:0em;">∑</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.9812em;"><span style="top:-2.4003em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mrel mtight">=</span><span class="mord mtight">1</span></span></span></span><span style="top:-3.2029em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight">G</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2997em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mord">∣</span><span class="mord"><span class="mord mathnormal">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight">i</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mord">∣</span></span></span><span style="top:-3.23em;"><span class="pstrut" style="height:3em;"></span><span class="frac-line" style="border-bottom-width:0.04em;"></span></span><span style="top:-3.677em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord">1</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:1.1709em;"><span></span></span></span></span></span><span class="mclose nulldelimiter"></span></span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mop op-limits"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.8283em;"><span style="top:-1.8723em;margin-left:0em;"><span class="pstrut" style="height:3.05em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mrel mtight">=</span><span class="mord mtight">1</span></span></span></span><span style="top:-3.05em;"><span class="pstrut" style="height:3.05em;"></span><span><span class="mop op-symbol large-op">∑</span></span></span><span style="top:-4.3em;margin-left:0em;"><span class="pstrut" style="height:3.05em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight">G</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:1.2777em;"><span></span></span></span></span></span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mop op-limits"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.961em;"><span style="top:-1.8829em;margin-left:0em;"><span class="pstrut" style="height:3.05em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">t</span><span class="mrel mtight">=</span><span class="mord mtight">1</span></span></span></span><span style="top:-3.05em;"><span class="pstrut" style="height:3.05em;"></span><span><span class="mop op-symbol large-op">∑</span></span></span><span style="top:-4.386em;margin-left:0em;"><span class="pstrut" style="height:3.05em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mtight">∣</span><span class="mord mtight"><span class="mord mathnormal mtight">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3281em;"><span style="top:-2.357em;margin-left:0em;margin-right:0.0714em;"><span class="pstrut" style="height:2.5em;"></span><span class="sizing reset-size3 size1 mtight"><span class="mord mathnormal mtight">i</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.143em;"><span></span></span></span></span></span></span><span class="mord mtight">∣</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:1.2671em;"><span></span></span></span></span></span><span class="mspace" style="margin-right:0.1667em;"></span><span class="minner"><span class="mopen delimcenter" style="top:0em;"><span class="delimsizing size4">[</span></span><span class="mord"><span class="mopen nulldelimiter"></span><span class="mfrac"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.427em;"><span style="top:-2.314em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="minner"><span class="minner"><span class="mopen delimcenter" style="top:0em;">[</span><span class="mord"><span class="mord mathnormal" style="margin-right:0.03588em;">π</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3361em;"><span style="top:-2.55em;margin-left:-0.0359em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight" style="margin-right:0.02778em;">θ</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mopen">(</span><span class="mord"><span class="mord mathnormal">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">∣</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mord mathnormal" style="margin-right:0.03588em;">q</span><span class="mpunct">,</span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mord"><span class="mord mathnormal">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mrel mtight"><</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mclose">)</span><span class="mclose delimcenter" style="top:0em;">]</span></span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.1503em;"><span style="top:-2.3642em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord text mtight"><span class="mord mtight">no grad</span></span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.4719em;"><span></span></span></span></span></span></span></span></span><span style="top:-3.23em;"><span class="pstrut" style="height:3em;"></span><span class="frac-line" style="border-bottom-width:0.04em;"></span></span><span style="top:-3.677em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord"><span class="mord mathnormal" style="margin-right:0.03588em;">π</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3361em;"><span style="top:-2.55em;margin-left:-0.0359em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight" style="margin-right:0.02778em;">θ</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mopen">(</span><span class="mord"><span class="mord mathnormal">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">∣</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mord mathnormal" style="margin-right:0.03588em;">q</span><span class="mpunct">,</span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mord"><span class="mord mathnormal">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mrel mtight"><</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mclose">)</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:1.1579em;"><span></span></span></span></span></span><span class="mclose nulldelimiter"></span></span><span class="mord"><span class="mord accent"><span class="vlist-t"><span class="vlist-r"><span class="vlist" style="height:0.9468em;"><span style="top:-3em;"><span class="pstrut" style="height:3em;"></span><span class="mord mathnormal">A</span></span><span style="top:-3.2523em;"><span class="pstrut" style="height:3em;"></span><span class="accent-body" style="left:-0.1111em;"><span class="mord">^</span></span></span></span></span></span></span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mbin">−</span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mord mathnormal" style="margin-right:0.05278em;">β</span><span class="mord"><span class="mord mathbb">D</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3283em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord text mtight"><span class="mord mtight">KL</span></span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.1667em;"></span><span class="minner"><span class="mopen delimcenter" style="top:0em;">[</span><span class="mord"><span class="mord mathnormal" style="margin-right:0.03588em;">π</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3361em;"><span style="top:-2.55em;margin-left:-0.0359em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight" style="margin-right:0.02778em;">θ</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mord">∥</span><span class="mord"><span class="mord mathnormal" style="margin-right:0.03588em;">π</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3361em;"><span style="top:-2.55em;margin-left:-0.0359em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord text mtight"><span class="mord mtight">ref</span></span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mclose delimcenter" style="top:0em;">]</span></span><span class="mclose delimcenter" style="top:0em;"><span class="delimsizing size4">]</span></span></span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mpunct">,</span></span></span></span></span><!-- HTML_TAG_END --></p> <p data-svelte-h="svelte-lefg4k">where the first term represents the scaled advantage and the second term penalizes deviations from the reference policy through KL divergence.</p> <blockquote class="tip"><p>Note that compared to the original formulation in <a href="https://huggingface.co/papers/2402.03300" rel="nofollow" data-svelte-h="svelte-16brarf">DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models</a>, we don’t scale by <!-- HTML_TAG_START --><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mfrac><mn>1</mn><mrow><mi mathvariant="normal">∣</mi><msub><mi>o</mi><mi>i</mi></msub><mi mathvariant="normal">∣</mi></mrow></mfrac></mrow><annotation encoding="application/x-tex"> \frac{1}{|o_i|} </annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:1.3651em;vertical-align:-0.52em;"></span><span class="mord"><span class="mopen nulldelimiter"></span><span class="mfrac"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.8451em;"><span style="top:-2.655em;"><span class="pstrut" style="height:3em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mtight">∣</span><span class="mord mtight"><span class="mord mathnormal mtight">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3281em;"><span style="top:-2.357em;margin-left:0em;margin-right:0.0714em;"><span class="pstrut" style="height:2.5em;"></span><span class="sizing reset-size3 size1 mtight"><span class="mord mathnormal mtight">i</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.143em;"><span></span></span></span></span></span></span><span class="mord mtight">∣</span></span></span></span><span style="top:-3.23em;"><span class="pstrut" style="height:3em;"></span><span class="frac-line" style="border-bottom-width:0.04em;"></span></span><span style="top:-3.394em;"><span class="pstrut" style="height:3em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mtight">1</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.52em;"><span></span></span></span></span></span><span class="mclose nulldelimiter"></span></span></span></span></span><!-- HTML_TAG_END --> because it was shown in the paper <a href="https://huggingface.co/papers/2503.20783" rel="nofollow" data-svelte-h="svelte-u44a8w">Understanding R1-Zero-Like Training: A Critical Perspective</a> that this introduces a response-level length bias. More details in <a href="#loss-types" data-svelte-h="svelte-o61jf4">loss types</a>.</p></blockquote> <blockquote class="tip"><p>Note that compared to the original formulation in <a href="https://huggingface.co/papers/2402.03300" rel="nofollow" data-svelte-h="svelte-16brarf">DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models</a>, we use <!-- HTML_TAG_START --><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>β</mi><mo>=</mo><mn>0.0</mn></mrow><annotation encoding="application/x-tex"> \beta = 0.0 </annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.8889em;vertical-align:-0.1944em;"></span><span class="mord mathnormal" style="margin-right:0.05278em;">β</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:0.6444em;"></span><span class="mord">0.0</span></span></span></span><!-- HTML_TAG_END --> by default, meaning that the KL divergence term is not used. This choice is motivated by several recent studies (e.g., <a href="https://huggingface.co/papers/2503.24290" rel="nofollow" data-svelte-h="svelte-1tkhwi2">Open-Reasoner-Zero: An Open Source Approach to Scaling Up Reinforcement Learning on the Base Model</a>) which have shown that the KL divergence term is not essential for training with GRPO. As a result, it has become common practice to exclude it (e.g. <a href="https://huggingface.co/papers/2503.20783" rel="nofollow" data-svelte-h="svelte-u44a8w">Understanding R1-Zero-Like Training: A Critical Perspective</a>, <a href="https://huggingface.co/papers/2503.14476" rel="nofollow" data-svelte-h="svelte-1q851rp">DAPO: An Open-Source LLM Reinforcement Learning System at Scale</a>). If you wish to include the KL divergence term, you can set <code data-svelte-h="svelte-ste5qq">beta</code> in <a href="/docs/trl/pr_4331/en/grpo_trainer#trl.GRPOConfig" data-svelte-h="svelte-jirtk">GRPOConfig</a> to a non-zero value.</p></blockquote> <p>In the original paper, this formulation is generalized to account for multiple updates after each generation (denoted <!-- HTML_TAG_START --><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>μ</mi></mrow><annotation encoding="application/x-tex"> \mu </annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.625em;vertical-align:-0.1944em;"></span><span class="mord mathnormal">μ</span></span></span></span><!-- HTML_TAG_END -->, can be set with <code data-svelte-h="svelte-1i6938t">num_iterations</code> in <a href="/docs/trl/pr_4331/en/grpo_trainer#trl.GRPOConfig" data-svelte-h="svelte-jirtk">GRPOConfig</a>) by leveraging the <strong data-svelte-h="svelte-18krrzk">clipped surrogate objective</strong>: | |
| <!-- HTML_TAG_START --><span class="katex-display"><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML" display="block"><semantics><mrow><msub><mi mathvariant="script">L</mi><mtext>GRPO</mtext></msub><mo stretchy="false">(</mo><mi>θ</mi><mo stretchy="false">)</mo><mo>=</mo><mo>−</mo><mfrac><mn>1</mn><mrow><munderover><mo>∑</mo><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><mi>G</mi></munderover><mi mathvariant="normal">∣</mi><msub><mi>o</mi><mi>i</mi></msub><mi mathvariant="normal">∣</mi></mrow></mfrac><munderover><mo>∑</mo><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><mi>G</mi></munderover><munderover><mo>∑</mo><mrow><mi>t</mi><mo>=</mo><mn>1</mn></mrow><mrow><mi mathvariant="normal">∣</mi><msub><mi>o</mi><mi>i</mi></msub><mi mathvariant="normal">∣</mi></mrow></munderover><mrow><mo fence="true">[</mo><mi>min</mi><mo></mo><mrow><mo fence="true">(</mo><mfrac><mrow><msub><mi>π</mi><mi>θ</mi></msub><mo stretchy="false">(</mo><msub><mi>o</mi><mrow><mi>i</mi><mo separator="true">,</mo><mi>t</mi></mrow></msub><mo>∣</mo><mi>q</mi><mo separator="true">,</mo><msub><mi>o</mi><mrow><mi>i</mi><mo separator="true">,</mo><mo><</mo><mi>t</mi></mrow></msub><mo stretchy="false">)</mo></mrow><mrow><msub><mi>π</mi><msub><mi>θ</mi><mtext>old</mtext></msub></msub><mo stretchy="false">(</mo><msub><mi>o</mi><mrow><mi>i</mi><mo separator="true">,</mo><mi>t</mi></mrow></msub><mo>∣</mo><mi>q</mi><mo separator="true">,</mo><msub><mi>o</mi><mrow><mi>i</mi><mo separator="true">,</mo><mo><</mo><mi>t</mi></mrow></msub><mo stretchy="false">)</mo></mrow></mfrac><msub><mover accent="true"><mi>A</mi><mo>^</mo></mover><mrow><mi>i</mi><mo separator="true">,</mo><mi>t</mi></mrow></msub><mo separator="true">,</mo><mtext> </mtext><mtext>clip</mtext><mrow><mo fence="true">(</mo><mfrac><mrow><msub><mi>π</mi><mi>θ</mi></msub><mo stretchy="false">(</mo><msub><mi>o</mi><mrow><mi>i</mi><mo separator="true">,</mo><mi>t</mi></mrow></msub><mo>∣</mo><mi>q</mi><mo separator="true">,</mo><msub><mi>o</mi><mrow><mi>i</mi><mo separator="true">,</mo><mo><</mo><mi>t</mi></mrow></msub><mo stretchy="false">)</mo></mrow><mrow><msub><mi>π</mi><msub><mi>θ</mi><mtext>old</mtext></msub></msub><mo stretchy="false">(</mo><msub><mi>o</mi><mrow><mi>i</mi><mo separator="true">,</mo><mi>t</mi></mrow></msub><mo>∣</mo><mi>q</mi><mo separator="true">,</mo><msub><mi>o</mi><mrow><mi>i</mi><mo separator="true">,</mo><mo><</mo><mi>t</mi></mrow></msub><mo stretchy="false">)</mo></mrow></mfrac><mo separator="true">,</mo><mn>1</mn><mo>−</mo><mi>ϵ</mi><mo separator="true">,</mo><mn>1</mn><mo>+</mo><mi>ϵ</mi><mo fence="true">)</mo></mrow><msub><mover accent="true"><mi>A</mi><mo>^</mo></mover><mrow><mi>i</mi><mo separator="true">,</mo><mi>t</mi></mrow></msub><mo fence="true">)</mo></mrow><mo>−</mo><mi>β</mi><msub><mi mathvariant="double-struck">D</mi><mtext>KL</mtext></msub><mrow><mo fence="true">[</mo><msub><mi>π</mi><mi>θ</mi></msub><mi mathvariant="normal">∥</mi><msub><mi>π</mi><mtext>ref</mtext></msub><mo fence="true">]</mo></mrow><mo fence="true">]</mo></mrow><mo separator="true">,</mo></mrow><annotation encoding="application/x-tex"> | |
| \mathcal{L}_{\text{GRPO}}(\theta) = - \frac{1}{\sum_{i=1}^G |o_i|} \sum_{i=1}^G \sum_{t=1}^{|o_i|} \left[ \min \left( \frac{\pi_\theta(o_{i,t} \mid q, o_{i,< t})}{\pi_{\theta_{\text{old}}}(o_{i,t} \mid q, o_{i,< t})} \hat{A}_{i,t}, \, \text{clip}\left( \frac{\pi_\theta(o_{i,t} \mid q, o_{i,< t})}{\pi_{\theta_{\text{old}}}(o_{i,t} \mid q, o_{i,< t})}, 1 - \epsilon, 1 + \epsilon \right) \hat{A}_{i,t} \right) - \beta \mathbb{D}_{\text{KL}}\left[\pi_\theta \| \pi_{\text{ref}}\right] \right], | |
| </annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord"><span class="mord mathcal">L</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3283em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord text mtight"><span class="mord mtight">GRPO</span></span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mopen">(</span><span class="mord mathnormal" style="margin-right:0.02778em;">θ</span><span class="mclose">)</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:3.2387em;vertical-align:-1.2777em;"></span><span class="mord">−</span><span class="mord"><span class="mopen nulldelimiter"></span><span class="mfrac"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.3214em;"><span style="top:-2.1288em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mop"><span class="mop op-symbol small-op" style="position:relative;top:0em;">∑</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.9812em;"><span style="top:-2.4003em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mrel mtight">=</span><span class="mord mtight">1</span></span></span></span><span style="top:-3.2029em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight">G</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2997em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mord">∣</span><span class="mord"><span class="mord mathnormal">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight">i</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mord">∣</span></span></span><span style="top:-3.23em;"><span class="pstrut" style="height:3em;"></span><span class="frac-line" style="border-bottom-width:0.04em;"></span></span><span style="top:-3.677em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord">1</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:1.1709em;"><span></span></span></span></span></span><span class="mclose nulldelimiter"></span></span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mop op-limits"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.8283em;"><span style="top:-1.8723em;margin-left:0em;"><span class="pstrut" style="height:3.05em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mrel mtight">=</span><span class="mord mtight">1</span></span></span></span><span style="top:-3.05em;"><span class="pstrut" style="height:3.05em;"></span><span><span class="mop op-symbol large-op">∑</span></span></span><span style="top:-4.3em;margin-left:0em;"><span class="pstrut" style="height:3.05em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight">G</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:1.2777em;"><span></span></span></span></span></span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mop op-limits"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.961em;"><span style="top:-1.8829em;margin-left:0em;"><span class="pstrut" style="height:3.05em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">t</span><span class="mrel mtight">=</span><span class="mord mtight">1</span></span></span></span><span style="top:-3.05em;"><span class="pstrut" style="height:3.05em;"></span><span><span class="mop op-symbol large-op">∑</span></span></span><span style="top:-4.386em;margin-left:0em;"><span class="pstrut" style="height:3.05em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mtight">∣</span><span class="mord mtight"><span class="mord mathnormal mtight">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3281em;"><span style="top:-2.357em;margin-left:0em;margin-right:0.0714em;"><span class="pstrut" style="height:2.5em;"></span><span class="sizing reset-size3 size1 mtight"><span class="mord mathnormal mtight">i</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.143em;"><span></span></span></span></span></span></span><span class="mord mtight">∣</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:1.2671em;"><span></span></span></span></span></span><span class="mspace" style="margin-right:0.1667em;"></span><span class="minner"><span class="mopen delimcenter" style="top:0em;"><span class="delimsizing size3">[</span></span><span class="mop">min</span><span class="mspace" style="margin-right:0.1667em;"></span><span class="minner"><span class="mopen delimcenter" style="top:0em;"><span class="delimsizing size3">(</span></span><span class="mord"><span class="mopen nulldelimiter"></span><span class="mfrac"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.427em;"><span style="top:-2.314em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord"><span class="mord mathnormal" style="margin-right:0.03588em;">π</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3361em;"><span style="top:-2.55em;margin-left:-0.0359em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mtight"><span class="mord mathnormal mtight" style="margin-right:0.02778em;">θ</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3448em;"><span style="top:-2.3488em;margin-left:-0.0278em;margin-right:0.0714em;"><span class="pstrut" style="height:2.5em;"></span><span class="sizing reset-size3 size1 mtight"><span class="mord mtight"><span class="mord text mtight"><span class="mord mtight">old</span></span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.1512em;"><span></span></span></span></span></span></span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2559em;"><span></span></span></span></span></span></span><span class="mopen">(</span><span class="mord"><span class="mord mathnormal">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">∣</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mord mathnormal" style="margin-right:0.03588em;">q</span><span class="mpunct">,</span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mord"><span class="mord mathnormal">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mrel mtight"><</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mclose">)</span></span></span><span style="top:-3.23em;"><span class="pstrut" style="height:3em;"></span><span class="frac-line" style="border-bottom-width:0.04em;"></span></span><span style="top:-3.677em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord"><span class="mord mathnormal" style="margin-right:0.03588em;">π</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3361em;"><span style="top:-2.55em;margin-left:-0.0359em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight" style="margin-right:0.02778em;">θ</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mopen">(</span><span class="mord"><span class="mord mathnormal">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">∣</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mord mathnormal" style="margin-right:0.03588em;">q</span><span class="mpunct">,</span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mord"><span class="mord mathnormal">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mrel mtight"><</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mclose">)</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.9721em;"><span></span></span></span></span></span><span class="mclose nulldelimiter"></span></span><span class="mord"><span class="mord accent"><span class="vlist-t"><span class="vlist-r"><span class="vlist" style="height:0.9468em;"><span style="top:-3em;"><span class="pstrut" style="height:3em;"></span><span class="mord mathnormal">A</span></span><span style="top:-3.2523em;"><span class="pstrut" style="height:3em;"></span><span class="accent-body" style="left:-0.1111em;"><span class="mord">^</span></span></span></span></span></span></span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mpunct">,</span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mord text"><span class="mord">clip</span></span><span class="mspace" style="margin-right:0.1667em;"></span><span class="minner"><span class="mopen delimcenter" style="top:0em;"><span class="delimsizing size3">(</span></span><span class="mord"><span class="mopen nulldelimiter"></span><span class="mfrac"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.427em;"><span style="top:-2.314em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord"><span class="mord mathnormal" style="margin-right:0.03588em;">π</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3361em;"><span style="top:-2.55em;margin-left:-0.0359em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mtight"><span class="mord mathnormal mtight" style="margin-right:0.02778em;">θ</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3448em;"><span style="top:-2.3488em;margin-left:-0.0278em;margin-right:0.0714em;"><span class="pstrut" style="height:2.5em;"></span><span class="sizing reset-size3 size1 mtight"><span class="mord mtight"><span class="mord text mtight"><span class="mord mtight">old</span></span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.1512em;"><span></span></span></span></span></span></span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2559em;"><span></span></span></span></span></span></span><span class="mopen">(</span><span class="mord"><span class="mord mathnormal">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">∣</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mord mathnormal" style="margin-right:0.03588em;">q</span><span class="mpunct">,</span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mord"><span class="mord mathnormal">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mrel mtight"><</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mclose">)</span></span></span><span style="top:-3.23em;"><span class="pstrut" style="height:3em;"></span><span class="frac-line" style="border-bottom-width:0.04em;"></span></span><span style="top:-3.677em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord"><span class="mord mathnormal" style="margin-right:0.03588em;">π</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3361em;"><span style="top:-2.55em;margin-left:-0.0359em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight" style="margin-right:0.02778em;">θ</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mopen">(</span><span class="mord"><span class="mord mathnormal">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">∣</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mord mathnormal" style="margin-right:0.03588em;">q</span><span class="mpunct">,</span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mord"><span class="mord mathnormal">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mrel mtight"><</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mclose">)</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.9721em;"><span></span></span></span></span></span><span class="mclose nulldelimiter"></span></span><span class="mpunct">,</span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mord">1</span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mbin">−</span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mord mathnormal">ϵ</span><span class="mpunct">,</span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mord">1</span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mbin">+</span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mord mathnormal">ϵ</span><span class="mclose delimcenter" style="top:0em;"><span class="delimsizing size3">)</span></span></span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mord"><span class="mord accent"><span class="vlist-t"><span class="vlist-r"><span class="vlist" style="height:0.9468em;"><span style="top:-3em;"><span class="pstrut" style="height:3em;"></span><span class="mord mathnormal">A</span></span><span style="top:-3.2523em;"><span class="pstrut" style="height:3em;"></span><span class="accent-body" style="left:-0.1111em;"><span class="mord">^</span></span></span></span></span></span></span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mclose delimcenter" style="top:0em;"><span class="delimsizing size3">)</span></span></span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mbin">−</span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mord mathnormal" style="margin-right:0.05278em;">β</span><span class="mord"><span class="mord mathbb">D</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3283em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord text mtight"><span class="mord mtight">KL</span></span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.1667em;"></span><span class="minner"><span class="mopen delimcenter" style="top:0em;">[</span><span class="mord"><span class="mord mathnormal" style="margin-right:0.03588em;">π</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3361em;"><span style="top:-2.55em;margin-left:-0.0359em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight" style="margin-right:0.02778em;">θ</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mord">∥</span><span class="mord"><span class="mord mathnormal" style="margin-right:0.03588em;">π</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3361em;"><span style="top:-2.55em;margin-left:-0.0359em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord text mtight"><span class="mord mtight">ref</span></span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mclose delimcenter" style="top:0em;">]</span></span><span class="mclose delimcenter" style="top:0em;"><span class="delimsizing size3">]</span></span></span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mpunct">,</span></span></span></span></span><!-- HTML_TAG_END --></p> <p>where <!-- HTML_TAG_START --><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mtext>clip</mtext><mo stretchy="false">(</mo><mo>⋅</mo><mo separator="true">,</mo><mn>1</mn><mo>−</mo><mi>ϵ</mi><mo separator="true">,</mo><mn>1</mn><mo>+</mo><mi>ϵ</mi><mo stretchy="false">)</mo></mrow><annotation encoding="application/x-tex">\text{clip}(\cdot, 1 - \epsilon, 1 + \epsilon) </annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord text"><span class="mord">clip</span></span><span class="mopen">(</span><span class="mord">⋅</span><span class="mpunct">,</span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mord">1</span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mbin">−</span><span class="mspace" style="margin-right:0.2222em;"></span></span><span class="base"><span class="strut" style="height:0.8389em;vertical-align:-0.1944em;"></span><span class="mord mathnormal">ϵ</span><span class="mpunct">,</span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mord">1</span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mbin">+</span><span class="mspace" style="margin-right:0.2222em;"></span></span><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord mathnormal">ϵ</span><span class="mclose">)</span></span></span></span><!-- HTML_TAG_END --> ensures that updates do not deviate excessively from the reference policy by bounding the policy ratio between <!-- HTML_TAG_START --><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mn>1</mn><mo>−</mo><mi>ϵ</mi></mrow><annotation encoding="application/x-tex"> 1 - \epsilon </annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.7278em;vertical-align:-0.0833em;"></span><span class="mord">1</span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mbin">−</span><span class="mspace" style="margin-right:0.2222em;"></span></span><span class="base"><span class="strut" style="height:0.4306em;"></span><span class="mord mathnormal">ϵ</span></span></span></span><!-- HTML_TAG_END --> and <!-- HTML_TAG_START --><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mn>1</mn><mo>+</mo><mi>ϵ</mi></mrow><annotation encoding="application/x-tex"> 1 + \epsilon </annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.7278em;vertical-align:-0.0833em;"></span><span class="mord">1</span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mbin">+</span><span class="mspace" style="margin-right:0.2222em;"></span></span><span class="base"><span class="strut" style="height:0.4306em;"></span><span class="mord mathnormal">ϵ</span></span></span></span><!-- HTML_TAG_END -->. | |
| When <!-- HTML_TAG_START --><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>μ</mi><mo>=</mo><mn>1</mn></mrow><annotation encoding="application/x-tex"> \mu = 1 </annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.625em;vertical-align:-0.1944em;"></span><span class="mord mathnormal">μ</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:0.6444em;"></span><span class="mord">1</span></span></span></span><!-- HTML_TAG_END --> (default in TRL), the clipped surrogate objective simplifies to the original objective.</p> <h4 class="relative group"><a id="loss-types" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#loss-types"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Loss Types</span></h4> <p>Several formulations of the objective have been proposed in the literature. Initially, the objective of GRPO was defined as follows: | |
| <!-- HTML_TAG_START --><span class="katex-display"><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML" display="block"><semantics><mrow><msub><mi mathvariant="script">L</mi><mtext>GRPO</mtext></msub><mo stretchy="false">(</mo><mi>θ</mi><mo stretchy="false">)</mo><mo>=</mo><mo>−</mo><mfrac><mn>1</mn><mi>G</mi></mfrac><munderover><mo>∑</mo><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><mi>G</mi></munderover><mfrac><mn>1</mn><mrow><mi mathvariant="normal">∣</mi><msub><mi>o</mi><mi>i</mi></msub><mi mathvariant="normal">∣</mi></mrow></mfrac><munderover><mo>∑</mo><mrow><mi>t</mi><mo>=</mo><mn>1</mn></mrow><mrow><mi mathvariant="normal">∣</mi><msub><mi>o</mi><mi>i</mi></msub><mi mathvariant="normal">∣</mi></mrow></munderover><msub><mi>l</mi><mrow><mi>i</mi><mo separator="true">,</mo><mi>t</mi></mrow></msub><mo separator="true">,</mo></mrow><annotation encoding="application/x-tex"> | |
| \mathcal{L}_{\text{GRPO}}(\theta) = - \frac{1}{G} \sum_{i=1}^G \frac{1}{|o_i|} \sum_{t=1}^{|o_i|} l_{i,t}, | |
| </annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord"><span class="mord mathcal">L</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3283em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord text mtight"><span class="mord mtight">GRPO</span></span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mopen">(</span><span class="mord mathnormal" style="margin-right:0.02778em;">θ</span><span class="mclose">)</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:3.2387em;vertical-align:-1.2777em;"></span><span class="mord">−</span><span class="mord"><span class="mopen nulldelimiter"></span><span class="mfrac"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.3214em;"><span style="top:-2.314em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord mathnormal">G</span></span></span><span style="top:-3.23em;"><span class="pstrut" style="height:3em;"></span><span class="frac-line" style="border-bottom-width:0.04em;"></span></span><span style="top:-3.677em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord">1</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.686em;"><span></span></span></span></span></span><span class="mclose nulldelimiter"></span></span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mop op-limits"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.8283em;"><span style="top:-1.8723em;margin-left:0em;"><span class="pstrut" style="height:3.05em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mrel mtight">=</span><span class="mord mtight">1</span></span></span></span><span style="top:-3.05em;"><span class="pstrut" style="height:3.05em;"></span><span><span class="mop op-symbol large-op">∑</span></span></span><span style="top:-4.3em;margin-left:0em;"><span class="pstrut" style="height:3.05em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight">G</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:1.2777em;"><span></span></span></span></span></span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mord"><span class="mopen nulldelimiter"></span><span class="mfrac"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.3214em;"><span style="top:-2.314em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord">∣</span><span class="mord"><span class="mord mathnormal">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight">i</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mord">∣</span></span></span><span style="top:-3.23em;"><span class="pstrut" style="height:3em;"></span><span class="frac-line" style="border-bottom-width:0.04em;"></span></span><span style="top:-3.677em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord">1</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.936em;"><span></span></span></span></span></span><span class="mclose nulldelimiter"></span></span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mop op-limits"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.961em;"><span style="top:-1.8829em;margin-left:0em;"><span class="pstrut" style="height:3.05em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">t</span><span class="mrel mtight">=</span><span class="mord mtight">1</span></span></span></span><span style="top:-3.05em;"><span class="pstrut" style="height:3.05em;"></span><span><span class="mop op-symbol large-op">∑</span></span></span><span style="top:-4.386em;margin-left:0em;"><span class="pstrut" style="height:3.05em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mtight">∣</span><span class="mord mtight"><span class="mord mathnormal mtight">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3281em;"><span style="top:-2.357em;margin-left:0em;margin-right:0.0714em;"><span class="pstrut" style="height:2.5em;"></span><span class="sizing reset-size3 size1 mtight"><span class="mord mathnormal mtight">i</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.143em;"><span></span></span></span></span></span></span><span class="mord mtight">∣</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:1.2671em;"><span></span></span></span></span></span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mord"><span class="mord mathnormal" style="margin-right:0.01968em;">l</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:-0.0197em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mpunct">,</span></span></span></span></span><!-- HTML_TAG_END --></p> <p>where | |
| <!-- HTML_TAG_START --><span class="katex-display"><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML" display="block"><semantics><mrow><msub><mi>l</mi><mrow><mi>i</mi><mo separator="true">,</mo><mi>t</mi></mrow></msub><mo>=</mo><mfrac><mrow><msub><mi>π</mi><mi>θ</mi></msub><mo stretchy="false">(</mo><msub><mi>o</mi><mrow><mi>i</mi><mo separator="true">,</mo><mi>t</mi></mrow></msub><mo>∣</mo><mi>q</mi><mo separator="true">,</mo><msub><mi>o</mi><mrow><mi>i</mi><mo separator="true">,</mo><mo><</mo><mi>t</mi></mrow></msub><mo stretchy="false">)</mo></mrow><msub><mrow><mo fence="true">[</mo><msub><mi>π</mi><mi>θ</mi></msub><mo stretchy="false">(</mo><msub><mi>o</mi><mrow><mi>i</mi><mo separator="true">,</mo><mi>t</mi></mrow></msub><mo>∣</mo><mi>q</mi><mo separator="true">,</mo><msub><mi>o</mi><mrow><mi>i</mi><mo separator="true">,</mo><mo><</mo><mi>t</mi></mrow></msub><mo stretchy="false">)</mo><mo fence="true">]</mo></mrow><mtext>no grad</mtext></msub></mfrac><msub><mover accent="true"><mi>A</mi><mo>^</mo></mover><mrow><mi>i</mi><mo separator="true">,</mo><mi>t</mi></mrow></msub><mo>−</mo><mi>β</mi><msub><mi mathvariant="double-struck">D</mi><mtext>KL</mtext></msub><mrow><mo fence="true">[</mo><msub><mi>π</mi><mi>θ</mi></msub><mi mathvariant="normal">∥</mi><msub><mi>π</mi><mtext>ref</mtext></msub><mo fence="true">]</mo></mrow><mi mathvariant="normal">.</mi></mrow><annotation encoding="application/x-tex"> | |
| l_{i,t} = \frac{\pi_\theta(o_{i,t} \mid q, o_{i,< t})}{\left[\pi_\theta(o_{i,t} \mid q, o_{i,< t})\right]_{\text{no grad}}} \hat{A}_{i,t} - \beta \mathbb{D}_{\text{KL}}\left[\pi_\theta \| \pi_{\text{ref}}\right]. | |
| </annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.9805em;vertical-align:-0.2861em;"></span><span class="mord"><span class="mord mathnormal" style="margin-right:0.01968em;">l</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:-0.0197em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:2.5849em;vertical-align:-1.1579em;"></span><span class="mord"><span class="mopen nulldelimiter"></span><span class="mfrac"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.427em;"><span style="top:-2.314em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="minner"><span class="minner"><span class="mopen delimcenter" style="top:0em;">[</span><span class="mord"><span class="mord mathnormal" style="margin-right:0.03588em;">π</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3361em;"><span style="top:-2.55em;margin-left:-0.0359em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight" style="margin-right:0.02778em;">θ</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mopen">(</span><span class="mord"><span class="mord mathnormal">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">∣</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mord mathnormal" style="margin-right:0.03588em;">q</span><span class="mpunct">,</span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mord"><span class="mord mathnormal">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mrel mtight"><</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mclose">)</span><span class="mclose delimcenter" style="top:0em;">]</span></span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.1503em;"><span style="top:-2.3642em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord text mtight"><span class="mord mtight">no grad</span></span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.4719em;"><span></span></span></span></span></span></span></span></span><span style="top:-3.23em;"><span class="pstrut" style="height:3em;"></span><span class="frac-line" style="border-bottom-width:0.04em;"></span></span><span style="top:-3.677em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord"><span class="mord mathnormal" style="margin-right:0.03588em;">π</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3361em;"><span style="top:-2.55em;margin-left:-0.0359em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight" style="margin-right:0.02778em;">θ</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mopen">(</span><span class="mord"><span class="mord mathnormal">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">∣</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mord mathnormal" style="margin-right:0.03588em;">q</span><span class="mpunct">,</span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mord"><span class="mord mathnormal">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mrel mtight"><</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mclose">)</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:1.1579em;"><span></span></span></span></span></span><span class="mclose nulldelimiter"></span></span><span class="mord"><span class="mord accent"><span class="vlist-t"><span class="vlist-r"><span class="vlist" style="height:0.9468em;"><span style="top:-3em;"><span class="pstrut" style="height:3em;"></span><span class="mord mathnormal">A</span></span><span style="top:-3.2523em;"><span class="pstrut" style="height:3em;"></span><span class="accent-body" style="left:-0.1111em;"><span class="mord">^</span></span></span></span></span></span></span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mbin">−</span><span class="mspace" style="margin-right:0.2222em;"></span></span><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord mathnormal" style="margin-right:0.05278em;">β</span><span class="mord"><span class="mord mathbb">D</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3283em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord text mtight"><span class="mord mtight">KL</span></span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.1667em;"></span><span class="minner"><span class="mopen delimcenter" style="top:0em;">[</span><span class="mord"><span class="mord mathnormal" style="margin-right:0.03588em;">π</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3361em;"><span style="top:-2.55em;margin-left:-0.0359em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight" style="margin-right:0.02778em;">θ</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mord">∥</span><span class="mord"><span class="mord mathnormal" style="margin-right:0.03588em;">π</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3361em;"><span style="top:-2.55em;margin-left:-0.0359em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord text mtight"><span class="mord mtight">ref</span></span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mclose delimcenter" style="top:0em;">]</span></span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mord">.</span></span></span></span></span><!-- HTML_TAG_END --></p> <p>The <a href="https://huggingface.co/papers/2503.14476" rel="nofollow" data-svelte-h="svelte-oig0q8">DAPO paper</a> highlights the limitations of the GRPO algorithm’s sample-level loss in long-CoT scenarios, where longer responses are under-penalized, leading to poorer quality outputs. The proposed solution is a token-level normalization, which better handles longer sequences by assigning more balanced rewards to individual tokens, regardless of response length: | |
| <!-- HTML_TAG_START --><span class="katex-display"><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML" display="block"><semantics><mrow><msub><mi mathvariant="script">L</mi><mtext>DAPO</mtext></msub><mo stretchy="false">(</mo><mi>θ</mi><mo stretchy="false">)</mo><mo>=</mo><mo>−</mo><mfrac><mn>1</mn><mrow><munderover><mo>∑</mo><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><mi>G</mi></munderover><mi mathvariant="normal">∣</mi><msub><mi>o</mi><mi>i</mi></msub><mi mathvariant="normal">∣</mi></mrow></mfrac><munderover><mo>∑</mo><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><mi>G</mi></munderover><munderover><mo>∑</mo><mrow><mi>t</mi><mo>=</mo><mn>1</mn></mrow><mrow><mi mathvariant="normal">∣</mi><msub><mi>o</mi><mi>i</mi></msub><mi mathvariant="normal">∣</mi></mrow></munderover><msub><mi>l</mi><mrow><mi>i</mi><mo separator="true">,</mo><mi>t</mi></mrow></msub><mo separator="true">,</mo></mrow><annotation encoding="application/x-tex"> | |
| \mathcal{L}_{\text{DAPO}}(\theta) = - \frac{1}{\sum_{i=1}^G |o_i|} \sum_{i=1}^G \sum_{t=1}^{|o_i|} l_{i,t}, | |
| </annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord"><span class="mord mathcal">L</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3283em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord text mtight"><span class="mord mtight">DAPO</span></span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mopen">(</span><span class="mord mathnormal" style="margin-right:0.02778em;">θ</span><span class="mclose">)</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:3.2387em;vertical-align:-1.2777em;"></span><span class="mord">−</span><span class="mord"><span class="mopen nulldelimiter"></span><span class="mfrac"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.3214em;"><span style="top:-2.1288em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mop"><span class="mop op-symbol small-op" style="position:relative;top:0em;">∑</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.9812em;"><span style="top:-2.4003em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mrel mtight">=</span><span class="mord mtight">1</span></span></span></span><span style="top:-3.2029em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight">G</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2997em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mord">∣</span><span class="mord"><span class="mord mathnormal">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight">i</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mord">∣</span></span></span><span style="top:-3.23em;"><span class="pstrut" style="height:3em;"></span><span class="frac-line" style="border-bottom-width:0.04em;"></span></span><span style="top:-3.677em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord">1</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:1.1709em;"><span></span></span></span></span></span><span class="mclose nulldelimiter"></span></span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mop op-limits"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.8283em;"><span style="top:-1.8723em;margin-left:0em;"><span class="pstrut" style="height:3.05em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mrel mtight">=</span><span class="mord mtight">1</span></span></span></span><span style="top:-3.05em;"><span class="pstrut" style="height:3.05em;"></span><span><span class="mop op-symbol large-op">∑</span></span></span><span style="top:-4.3em;margin-left:0em;"><span class="pstrut" style="height:3.05em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight">G</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:1.2777em;"><span></span></span></span></span></span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mop op-limits"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.961em;"><span style="top:-1.8829em;margin-left:0em;"><span class="pstrut" style="height:3.05em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">t</span><span class="mrel mtight">=</span><span class="mord mtight">1</span></span></span></span><span style="top:-3.05em;"><span class="pstrut" style="height:3.05em;"></span><span><span class="mop op-symbol large-op">∑</span></span></span><span style="top:-4.386em;margin-left:0em;"><span class="pstrut" style="height:3.05em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mtight">∣</span><span class="mord mtight"><span class="mord mathnormal mtight">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3281em;"><span style="top:-2.357em;margin-left:0em;margin-right:0.0714em;"><span class="pstrut" style="height:2.5em;"></span><span class="sizing reset-size3 size1 mtight"><span class="mord mathnormal mtight">i</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.143em;"><span></span></span></span></span></span></span><span class="mord mtight">∣</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:1.2671em;"><span></span></span></span></span></span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mord"><span class="mord mathnormal" style="margin-right:0.01968em;">l</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:-0.0197em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mpunct">,</span></span></span></span></span><!-- HTML_TAG_END --></p> <p data-svelte-h="svelte-qdorxk">To use this formulation, set <code>loss_type="dapo"</code> in <a href="/docs/trl/pr_4331/en/grpo_trainer#trl.GRPOConfig">GRPOConfig</a>.</p> <p>Furthermore, it was demonstrated in the paper <a href="https://huggingface.co/papers/2503.20783" rel="nofollow" data-svelte-h="svelte-u44a8w">Understanding R1-Zero-Like Training: A Critical Perspective</a> that the initial GRPO formulation introduces a response length bias. They show that while the DAPO formulation reduces this bias, it does not eliminate it completely. To fully remove this bias, they propose dividing by a constant instead of the sequence length, resulting in the following formulation: | |
| <!-- HTML_TAG_START --><span class="katex-display"><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML" display="block"><semantics><mrow><msub><mi mathvariant="script">L</mi><mtext>Dr. GRPO</mtext></msub><mo stretchy="false">(</mo><mi>θ</mi><mo stretchy="false">)</mo><mo>=</mo><mo>−</mo><mfrac><mn>1</mn><mrow><mi>L</mi><mi>G</mi></mrow></mfrac><munderover><mo>∑</mo><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><mi>G</mi></munderover><munderover><mo>∑</mo><mrow><mi>t</mi><mo>=</mo><mn>1</mn></mrow><mrow><mi mathvariant="normal">∣</mi><msub><mi>o</mi><mi>i</mi></msub><mi mathvariant="normal">∣</mi></mrow></munderover><msub><mi>l</mi><mrow><mi>i</mi><mo separator="true">,</mo><mi>t</mi></mrow></msub><mo separator="true">,</mo></mrow><annotation encoding="application/x-tex"> | |
| \mathcal{L}_{\text{Dr. GRPO}}(\theta) = - \frac{1}{LG} \sum_{i=1}^G \sum_{t=1}^{|o_i|} l_{i,t}, | |
| </annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord"><span class="mord mathcal">L</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3283em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord text mtight"><span class="mord mtight">Dr. GRPO</span></span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mopen">(</span><span class="mord mathnormal" style="margin-right:0.02778em;">θ</span><span class="mclose">)</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:3.2387em;vertical-align:-1.2777em;"></span><span class="mord">−</span><span class="mord"><span class="mopen nulldelimiter"></span><span class="mfrac"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.3214em;"><span style="top:-2.314em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord mathnormal">L</span><span class="mord mathnormal">G</span></span></span><span style="top:-3.23em;"><span class="pstrut" style="height:3em;"></span><span class="frac-line" style="border-bottom-width:0.04em;"></span></span><span style="top:-3.677em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord">1</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.686em;"><span></span></span></span></span></span><span class="mclose nulldelimiter"></span></span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mop op-limits"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.8283em;"><span style="top:-1.8723em;margin-left:0em;"><span class="pstrut" style="height:3.05em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mrel mtight">=</span><span class="mord mtight">1</span></span></span></span><span style="top:-3.05em;"><span class="pstrut" style="height:3.05em;"></span><span><span class="mop op-symbol large-op">∑</span></span></span><span style="top:-4.3em;margin-left:0em;"><span class="pstrut" style="height:3.05em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight">G</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:1.2777em;"><span></span></span></span></span></span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mop op-limits"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.961em;"><span style="top:-1.8829em;margin-left:0em;"><span class="pstrut" style="height:3.05em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">t</span><span class="mrel mtight">=</span><span class="mord mtight">1</span></span></span></span><span style="top:-3.05em;"><span class="pstrut" style="height:3.05em;"></span><span><span class="mop op-symbol large-op">∑</span></span></span><span style="top:-4.386em;margin-left:0em;"><span class="pstrut" style="height:3.05em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mtight">∣</span><span class="mord mtight"><span class="mord mathnormal mtight">o</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3281em;"><span style="top:-2.357em;margin-left:0em;margin-right:0.0714em;"><span class="pstrut" style="height:2.5em;"></span><span class="sizing reset-size3 size1 mtight"><span class="mord mathnormal mtight">i</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.143em;"><span></span></span></span></span></span></span><span class="mord mtight">∣</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:1.2671em;"><span></span></span></span></span></span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mord"><span class="mord mathnormal" style="margin-right:0.01968em;">l</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:-0.0197em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mpunct">,</span></span></span></span></span><!-- HTML_TAG_END --></p> <p data-svelte-h="svelte-99yhng">This constant is recommended to be the maximum completion length. To use this formulation, set <code>loss_type="dr_grpo"</code> in the <a href="/docs/trl/pr_4331/en/grpo_trainer#trl.GRPOConfig">GRPOConfig</a>.</p> <h2 class="relative group"><a id="logged-metrics" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#logged-metrics"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Logged metrics</span></h2> <p data-svelte-h="svelte-vze52b">While training and evaluating, we record the following reward metrics:</p> <ul><li data-svelte-h="svelte-qewuc3"><code>num_tokens</code>: The total number of tokens processed so far, including both prompts and completions.</li> <li data-svelte-h="svelte-a0iv18"><code>completions/mean_length</code>: The average length of generated completions.</li> <li data-svelte-h="svelte-1kqjbl8"><code>completions/min_length</code>: The minimum length of generated completions.</li> <li data-svelte-h="svelte-1xrlq9c"><code>completions/max_length</code>: The maximum length of generated completions.</li> <li data-svelte-h="svelte-1x1525x"><code>completions/mean_terminated_length</code>: The average length of generated completions that terminate with EOS.</li> <li data-svelte-h="svelte-ef26ix"><code>completions/min_terminated_length</code>: The minimum length of generated completions that terminate with EOS.</li> <li data-svelte-h="svelte-1dkehpx"><code>completions/max_terminated_length</code>: The maximum length of generated completions that terminate with EOS.</li> <li data-svelte-h="svelte-dpkiu2"><code>completions/clipped_ratio</code>: The ratio of truncated (clipped) completions.</li> <li data-svelte-h="svelte-j7e6bo"><code>reward/{reward_func_name}/mean</code>: The average reward from a specific reward function.</li> <li data-svelte-h="svelte-scyo6b"><code>reward/{reward_func_name}/std</code>: The standard deviation of the reward from a specific reward function.</li> <li data-svelte-h="svelte-w4pemg"><code>reward</code>: The overall average reward after applying reward weights.</li> <li data-svelte-h="svelte-1nkqtth"><code>reward_std</code>: The standard deviation of rewards after applying reward weights. | |
| <ul><li>If <code>scale_rewards</code> is <code>"group"</code> or <code>"none"</code>, this is the average of the per-group standard deviations.</li> <li>If <code>scale_rewards</code> is <code>"batch"</code>, this is the standard deviation computed over all rewards in the batch (ignoring groups).</li></ul></li> <li data-svelte-h="svelte-1xvloak"><code>frac_reward_zero_std</code>: The fraction of samples in the generation batch with a reward std of zero, implying there is little diversity for that prompt (all answers are correct or incorrect).</li> <li data-svelte-h="svelte-153aat5"><code>entropy</code>: Average entropy of token predictions across generated completions. (If <code>mask_truncated_completions=True</code>, masked sequences tokens are excluded.)</li> <li data-svelte-h="svelte-k2atia"><code>kl</code>: The average KL divergence between the model and the reference model, calculated over generated completions. Logged only if <code>beta</code> is nonzero.</li> <li><code data-svelte-h="svelte-iqp3k9">clip_ratio/region_mean</code>: The ratio of token (or sequence, if <code data-svelte-h="svelte-d1rrl1">importance_sampling_level="sequence"</code>) probabilities where the GRPO objective is clipped to stay within the trust region: | |
| $$ | |
| \text{clip}\left( r<em data-svelte-h="svelte-1ba4qaf">{i,t}(\theta), 1 - \epsilon</em>\mathrm{low}, 1 + \epsilon<em data-svelte-h="svelte-2ok3dp">\mathrm{high} \right)\,, \qquad r</em>{i,t}(\theta) = \frac{\pi<em data-svelte-h="svelte-1djdnpz">\theta(o</em>{i,t} \mid q, o<em data-svelte-h="svelte-1m37xqt">{i,< t})}{\pi</em>{\theta<em data-svelte-h="svelte-qlg616">{\text{old}}}(o</em>{i,t} \mid q, o_{i,< t})}\,. | |
| $$ | |
| A higher value means more tokens are clipped, which constrains how much the policy <!-- HTML_TAG_START --><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mi>π</mi><mi>θ</mi></msub></mrow><annotation encoding="application/x-tex">\pi_\theta</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.5806em;vertical-align:-0.15em;"></span><span class="mord"><span class="mord mathnormal" style="margin-right:0.03588em;">π</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3361em;"><span style="top:-2.55em;margin-left:-0.0359em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight" style="margin-right:0.02778em;">θ</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span></span></span></span><!-- HTML_TAG_END --> can change.</li> <li><code data-svelte-h="svelte-1wl9jmh">clip_ratio/low_mean</code>: The average ratio of token (or sequence, if <code data-svelte-h="svelte-d1rrl1">importance_sampling_level="sequence"</code>) probabilities that were clipped on the lower bound of the trust region: <!-- HTML_TAG_START --><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mi>r</mi><mrow><mi>i</mi><mo separator="true">,</mo><mi>t</mi></mrow></msub><mo stretchy="false">(</mo><mi>θ</mi><mo stretchy="false">)</mo><mo><</mo><mn>1</mn><mo>−</mo><msub><mi>ϵ</mi><mrow><mi mathvariant="normal">l</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">w</mi></mrow></msub></mrow><annotation encoding="application/x-tex">r_{i,t}(\theta) < 1 - \epsilon_\mathrm{low}</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:1.0361em;vertical-align:-0.2861em;"></span><span class="mord"><span class="mord mathnormal" style="margin-right:0.02778em;">r</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:-0.0278em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mopen">(</span><span class="mord mathnormal" style="margin-right:0.02778em;">θ</span><span class="mclose">)</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel"><</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:0.7278em;vertical-align:-0.0833em;"></span><span class="mord">1</span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mbin">−</span><span class="mspace" style="margin-right:0.2222em;"></span></span><span class="base"><span class="strut" style="height:0.5806em;vertical-align:-0.15em;"></span><span class="mord"><span class="mord mathnormal">ϵ</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3361em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathrm mtight" style="margin-right:0.01389em;">low</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span></span></span></span><!-- HTML_TAG_END --></li> <li><code data-svelte-h="svelte-zggxjg">clip_ratio/low_min</code>: The minimum ratio of token (or sequence, if <code data-svelte-h="svelte-d1rrl1">importance_sampling_level="sequence"</code>) probabilities that were clipped on the lower bound of the trust region: <!-- HTML_TAG_START --><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mi>r</mi><mrow><mi>i</mi><mo separator="true">,</mo><mi>t</mi></mrow></msub><mo stretchy="false">(</mo><mi>θ</mi><mo stretchy="false">)</mo><mo><</mo><mn>1</mn><mo>−</mo><msub><mi>ϵ</mi><mrow><mi mathvariant="normal">l</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">w</mi></mrow></msub></mrow><annotation encoding="application/x-tex">r_{i,t}(\theta) < 1 - \epsilon_\mathrm{low}</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:1.0361em;vertical-align:-0.2861em;"></span><span class="mord"><span class="mord mathnormal" style="margin-right:0.02778em;">r</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:-0.0278em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mopen">(</span><span class="mord mathnormal" style="margin-right:0.02778em;">θ</span><span class="mclose">)</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel"><</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:0.7278em;vertical-align:-0.0833em;"></span><span class="mord">1</span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mbin">−</span><span class="mspace" style="margin-right:0.2222em;"></span></span><span class="base"><span class="strut" style="height:0.5806em;vertical-align:-0.15em;"></span><span class="mord"><span class="mord mathnormal">ϵ</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3361em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathrm mtight" style="margin-right:0.01389em;">low</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span></span></span></span><!-- HTML_TAG_END --></li> <li><code data-svelte-h="svelte-1tram1d">clip_ratio/high_mean</code>: The average ratio of token (or sequence, if <code data-svelte-h="svelte-d1rrl1">importance_sampling_level="sequence"</code>) probabilities that were clipped on the upper bound of the trust region: <!-- HTML_TAG_START --><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mi>r</mi><mrow><mi>i</mi><mo separator="true">,</mo><mi>t</mi></mrow></msub><mo stretchy="false">(</mo><mi>θ</mi><mo stretchy="false">)</mo><mo>></mo><mn>1</mn><mo>+</mo><msub><mi>ϵ</mi><mrow><mi mathvariant="normal">h</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">g</mi><mi mathvariant="normal">h</mi></mrow></msub></mrow><annotation encoding="application/x-tex">r_{i,t}(\theta) > 1 + \epsilon_\mathrm{high}</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:1.0361em;vertical-align:-0.2861em;"></span><span class="mord"><span class="mord mathnormal" style="margin-right:0.02778em;">r</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:-0.0278em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mopen">(</span><span class="mord mathnormal" style="margin-right:0.02778em;">θ</span><span class="mclose">)</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">></span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:0.7278em;vertical-align:-0.0833em;"></span><span class="mord">1</span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mbin">+</span><span class="mspace" style="margin-right:0.2222em;"></span></span><span class="base"><span class="strut" style="height:0.7167em;vertical-align:-0.2861em;"></span><span class="mord"><span class="mord mathnormal">ϵ</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3361em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathrm mtight">high</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span></span></span></span><!-- HTML_TAG_END --></li> <li><code data-svelte-h="svelte-g4lei0">clip_ratio/high_max</code>: The maximum ratio of token (or sequence, if <code data-svelte-h="svelte-d1rrl1">importance_sampling_level="sequence"</code>) probabilities that were clipped on the upper bound of the trust region: <!-- HTML_TAG_START --><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mi>r</mi><mrow><mi>i</mi><mo separator="true">,</mo><mi>t</mi></mrow></msub><mo stretchy="false">(</mo><mi>θ</mi><mo stretchy="false">)</mo><mo>></mo><mn>1</mn><mo>+</mo><msub><mi>ϵ</mi><mrow><mi mathvariant="normal">h</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">g</mi><mi mathvariant="normal">h</mi></mrow></msub></mrow><annotation encoding="application/x-tex">r_{i,t}(\theta) > 1 + \epsilon_\mathrm{high}</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:1.0361em;vertical-align:-0.2861em;"></span><span class="mord"><span class="mord mathnormal" style="margin-right:0.02778em;">r</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:-0.0278em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span><span class="mpunct mtight">,</span><span class="mord mathnormal mtight">t</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mopen">(</span><span class="mord mathnormal" style="margin-right:0.02778em;">θ</span><span class="mclose">)</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">></span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:0.7278em;vertical-align:-0.0833em;"></span><span class="mord">1</span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mbin">+</span><span class="mspace" style="margin-right:0.2222em;"></span></span><span class="base"><span class="strut" style="height:0.7167em;vertical-align:-0.2861em;"></span><span class="mord"><span class="mord mathnormal">ϵ</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3361em;"><span style="top:-2.55em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathrm mtight">high</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span></span></span></span><!-- HTML_TAG_END -->.</li></ul> <h2 class="relative group"><a id="customization" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#customization"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Customization</span></h2> <h3 class="relative group"><a id="speed-up-training-with-vllm-powered-generation" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#speed-up-training-with-vllm-powered-generation"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Speed up training with vLLM-powered generation</span></h3> <p data-svelte-h="svelte-uhtnkr">Generation is often the main bottleneck when training with online methods. To accelerate generation, you can use <a href="https://github.com/vllm-project/vllm" rel="nofollow">vLLM</a>, a high-throughput, low-latency inference engine for LLMs. To enable it, first install the package with</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->pip install trl[vllm]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-wpkh1u">We support two ways of using vLLM during training: <strong>server mode</strong> and <strong>colocate mode</strong>.</p> <blockquote class="tip" data-svelte-h="svelte-19brdw3"><p>By default, Truncated Importance Sampling is activated for vLLM generation to address the generation-training mismatch that occurs when using different frameworks. This can be turned off by setting <code>vllm_importance_sampling_correction=False</code>. For more information, see <a href="paper_index#truncated-importance-sampling">Truncated Importance Sampling</a></p></blockquote> <h4 class="relative group"><a id="-option-1-server-mode" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#-option-1-server-mode"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>🔌 Option 1: Server mode</span></h4> <p data-svelte-h="svelte-14qro2b">In this mode, vLLM runs in a separate process (and using separate GPUs) and communicates with the trainer via HTTP. This is ideal if you have dedicated GPUs for inference.</p> <ol><li><p data-svelte-h="svelte-c2qog"><strong>Start the vLLM server</strong>:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->trl vllm-serve --model <model_name><!-- HTML_TAG_END --></pre></div></li> <li><p data-svelte-h="svelte-l0zqox"><strong>Enable server mode in your training script</strong>:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> trl <span class="hljs-keyword">import</span> GRPOConfig | |
| training_args = GRPOConfig( | |
| ..., | |
| use_vllm=<span class="hljs-literal">True</span>, | |
| vllm_mode=<span class="hljs-string">"server"</span>, <span class="hljs-comment"># default value, can be omitted</span> | |
| )<!-- HTML_TAG_END --></pre></div></li></ol> <blockquote class="warning" data-svelte-h="svelte-16p5pwe"><p>Make sure that the server is using different GPUs than the trainer, otherwise you may run into NCCL errors. You can specify the GPUs to use with the <code>CUDA_VISIBLE_DEVICES</code> environment variable.</p></blockquote> <h4 class="relative group"><a id="-option-2-colocate-mode" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#-option-2-colocate-mode"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>🧩 Option 2: Colocate mode</span></h4> <p data-svelte-h="svelte-g8ygxn">In this mode, vLLM runs inside the trainer process and shares GPU memory with the training model. This avoids launching a separate server and can improve GPU utilization, but may lead to memory contention on the training GPUs.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> trl <span class="hljs-keyword">import</span> GRPOConfig | |
| training_args = GRPOConfig( | |
| ..., | |
| use_vllm=<span class="hljs-literal">True</span>, | |
| vllm_mode=<span class="hljs-string">"colocate"</span>, | |
| )<!-- HTML_TAG_END --></pre></div> <blockquote class="tip" data-svelte-h="svelte-w0sg0a"><p>Depending on the model size and the overall GPU memory requirements for training, you may need to adjust the <code>vllm_gpu_memory_utilization</code> parameter in <a href="/docs/trl/pr_4331/en/grpo_trainer#trl.GRPOConfig">GRPOConfig</a> to avoid underutilization or out-of-memory errors.</p> <p>We provide a <a href="https://huggingface.co/spaces/trl-lib/recommend-vllm-memory" rel="nofollow">HF Space</a> to help estimate the recommended GPU memory utilization based on your model configuration and experiment settings. Simply use it as follows to get <code>vllm_gpu_memory_utilization</code> recommendation:</p> <iframe src="https://trl-lib-recommend-vllm-memory.hf.space" frameborder="0" width="850" height="450"></iframe> <p>If the recommended value does not work in your environment, we suggest adding a small buffer (e.g., +0.05 or +0.1) to the recommended value to ensure stability.</p> <p>If you still find you are getting out-of-memory errors set <code>vllm_enable_sleep_mode</code> to True and the vllm parameters and cache will be offloaded during the optimization step. For more information, see <a href="reducing_memory_usage#vllm-sleep-mode">Reducing Memory Usage with vLLM Sleep Mode</a>.</p></blockquote> <blockquote class="tip" data-svelte-h="svelte-9yuadg"><p>By default, GRPO uses <code>MASTER_ADDR=localhost</code> and <code>MASTER_PORT=12345</code> for vLLM, but you can override these values by setting the environment variables accordingly.</p></blockquote> <p data-svelte-h="svelte-4esodc">For more information, see <a href="speeding_up_training#vllm-for-fast-generation-in-online-methods">Speeding up training with vLLM</a>.</p> <h3 class="relative group"><a id="grpo-at-scale-train-a-70b-model-on-multiple-nodes" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#grpo-at-scale-train-a-70b-model-on-multiple-nodes"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>GRPO at scale: train a 70B+ Model on multiple nodes</span></h3> <p data-svelte-h="svelte-1lkshlz">When training large models like <strong>Qwen2.5-72B</strong>, you need several key optimizations to make the training efficient and scalable across multiple GPUs and nodes. These include:</p> <ul data-svelte-h="svelte-1fpjj5p"><li><strong>DeepSpeed ZeRO Stage 3</strong>: ZeRO leverages data parallelism to distribute model states (weights, gradients, optimizer states) across multiple GPUs and CPUs, reducing memory and compute requirements on each device. Since large models cannot fit on a single GPU, using ZeRO Stage 3 is required for training such models. For more details, see <a href="deepspeed_integration">DeepSpeed Integration</a>.</li> <li><strong>Accelerate</strong>: Accelerate is a library that simplifies distributed training across multiple GPUs and nodes. It provides a simple API to launch distributed training and handles the complexities of distributed training, such as data parallelism, gradient accumulation, and distributed data loading. For more details, see <a href="distributing_training">Distributing Training</a>.</li> <li><strong>vLLM</strong>: See the previous section on how to use vLLM to speed up generation.</li></ul> <p data-svelte-h="svelte-1d00z2m">Below is an example SLURM script to train a 70B model with GRPO on multiple nodes. This script trains a model on 4 nodes and uses the 5th node for vLLM-powered generation.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-meta">#!/bin/bash</span> | |
| <span class="hljs-comment">#SBATCH --nodes=5</span> | |
| <span class="hljs-comment">#SBATCH --gres=gpu:8</span> | |
| <span class="hljs-comment"># Get the list of allocated nodes</span> | |
| NODELIST=($(scontrol show hostnames <span class="hljs-variable">$SLURM_JOB_NODELIST</span>)) | |
| <span class="hljs-comment"># Assign the first 4 nodes for training and the 5th node for vLLM</span> | |
| TRAIN_NODES=<span class="hljs-string">"<span class="hljs-variable">${NODELIST[@]:0:4}</span>"</span> <span class="hljs-comment"># Nodes 0, 1, 2, 3 for training</span> | |
| VLLM_NODE=<span class="hljs-string">"<span class="hljs-variable">${NODELIST[4]}</span>"</span> <span class="hljs-comment"># Node 4 for vLLM</span> | |
| <span class="hljs-comment"># Run training on the first 4 nodes (Group 1)</span> | |
| srun --nodes=4 --ntasks=4 --nodelist=<span class="hljs-string">"<span class="hljs-variable">${NODELIST[@]:0:4}</span>"</span> accelerate launch \ | |
| --config_file examples/accelerate_configs/deepspeed_zero3.yaml \ | |
| --num_processes 32 \ | |
| --num_machines 4 \ | |
| --main_process_ip <span class="hljs-variable">${NODELIST[0]}</span> \ | |
| --machine_rank <span class="hljs-variable">$SLURM_PROCID</span> \ | |
| --rdzv_backend c10d \ | |
| train_grpo.py \ | |
| --server_ip <span class="hljs-variable">$VLLM_NODE</span> & | |
| <span class="hljs-comment"># Run vLLM server on the 5th node (Group 2)</span> | |
| srun --nodes=1 --ntasks=1 --nodelist=<span class="hljs-string">"<span class="hljs-variable">${NODELIST[4]}</span>"</span> trl vllm-serve --model Qwen/Qwen2.5-72B --tensor_parallel_size 8 & | |
| <span class="hljs-built_in">wait</span><!-- HTML_TAG_END --></pre></div> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">import</span> argparse | |
| <span class="hljs-keyword">from</span> datasets <span class="hljs-keyword">import</span> load_dataset | |
| <span class="hljs-keyword">from</span> trl <span class="hljs-keyword">import</span> GRPOTrainer, GRPOConfig | |
| <span class="hljs-keyword">def</span> <span class="hljs-title function_">main</span>(): | |
| parser = argparse.ArgumentParser() | |
| parser.add_argument(<span class="hljs-string">"--vllm_server_host"</span>, <span class="hljs-built_in">type</span>=<span class="hljs-built_in">str</span>, default=<span class="hljs-string">""</span>, <span class="hljs-built_in">help</span>=<span class="hljs-string">"The server IP"</span>) | |
| args = parser.parse_args() | |
| <span class="hljs-comment"># Example dataset from TLDR</span> | |
| dataset = load_dataset(<span class="hljs-string">"trl-lib/tldr"</span>, split=<span class="hljs-string">"train"</span>) | |
| <span class="hljs-comment"># Dummy reward function: count the number of unique characters in the completions</span> | |
| <span class="hljs-keyword">def</span> <span class="hljs-title function_">reward_num_unique_chars</span>(<span class="hljs-params">completions, **kwargs</span>): | |
| <span class="hljs-keyword">return</span> [<span class="hljs-built_in">len</span>(<span class="hljs-built_in">set</span>(c)) <span class="hljs-keyword">for</span> c <span class="hljs-keyword">in</span> completions] | |
| training_args = GRPOConfig( | |
| output_dir=<span class="hljs-string">"Qwen2.5-72B-GRPO"</span>, | |
| per_device_train_batch_size=<span class="hljs-number">4</span>, | |
| bf16=<span class="hljs-literal">True</span>, | |
| gradient_checkpointing=<span class="hljs-literal">True</span>, | |
| use_vllm=<span class="hljs-literal">True</span>, | |
| vllm_server_host=args.vllm_server_host.replace(<span class="hljs-string">"ip-"</span>, <span class="hljs-string">""</span>).replace(<span class="hljs-string">"-"</span>, <span class="hljs-string">"."</span>), <span class="hljs-comment"># from ip-X-X-X-X to X.X.X.X</span> | |
| ) | |
| trainer = GRPOTrainer(model=<span class="hljs-string">"Qwen/Qwen2.5-72B"</span>, args=training_args, reward_funcs=reward_num_unique_chars, train_dataset=dataset) | |
| trainer.train() | |
| <span class="hljs-keyword">if</span> __name__==<span class="hljs-string">"__main__"</span>: | |
| main()<!-- HTML_TAG_END --></pre></div> <h3 class="relative group"><a id="using-a-custom-reward-function" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#using-a-custom-reward-function"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Using a custom reward function</span></h3> <p data-svelte-h="svelte-13yb01z">The <a href="/docs/trl/pr_4331/en/grpo_trainer#trl.GRPOTrainer">GRPOTrainer</a> supports using custom reward functions instead of dense reward models. To ensure compatibility, your reward function must satisfy the following requirements:</p> <ol data-svelte-h="svelte-1m3leey"><li><p><strong>Input arguments</strong>:</p> <ul><li><p>The function must accept the following as keyword arguments:</p> <ul><li><code>prompts</code> (contains the prompts),</li> <li><code>completions</code> (contains the generated completions),</li> <li><code>completions_ids</code> (contains the tokenized completions),</li> <li><code>trainer_state</code> (<a href="https://huggingface.co/docs/transformers/main/en/main_classes/callback#transformers.TrainerState" rel="nofollow">TrainerState</a>): The current state of the trainer. This can be used to implement dynamic reward functions, such as curriculum learning, where the reward is adjusted based on the training progress.</li> <li>All column names (but <code>prompt</code>) that the dataset may have. For example, if the dataset contains a column named <code>ground_truth</code>, the function will be called with <code>ground_truth</code> as a keyword argument.</li></ul> <p>The easiest way to comply with this requirement is to use <code>**kwargs</code> in the function signature.</p></li> <li><p>Depending on the dataset format, the input will vary:</p> <ul><li>For <a href="dataset_formats#standard">standard format</a>, <code>prompts</code> and <code>completions</code> will be lists of strings.</li> <li>For <a href="dataset_formats#conversational">conversational format</a>, <code>prompts</code> and <code>completions</code> will be lists of message dictionaries.</li></ul></li></ul></li> <li><p><strong>Return value</strong>: The function must return a list of floats. Each float represents the reward corresponding to a single completion.</p></li></ol> <h4 class="relative group"><a id="example-1-reward-longer-completions" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#example-1-reward-longer-completions"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Example 1: Reward longer completions</span></h4> <p data-svelte-h="svelte-1n7vn3s">Below is an example of a reward function for a standard format that rewards longer completions:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">def</span> <span class="hljs-title function_">reward_func</span>(<span class="hljs-params">completions_ids, **kwargs</span>): | |
| <span class="hljs-string">"""Reward function that assigns higher scores to longer completions (in terms of token count)."""</span> | |
| <span class="hljs-keyword">return</span> [<span class="hljs-built_in">float</span>(<span class="hljs-built_in">len</span>(ids)) <span class="hljs-keyword">for</span> ids <span class="hljs-keyword">in</span> completions_ids]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-17jijhc">You can test it as follows:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>prompts = [<span class="hljs-string">"The sky is"</span>, <span class="hljs-string">"The sun is"</span>] <span class="hljs-comment"># not used in the reward function, but the trainer will pass it</span> | |
| <span class="hljs-meta">>>> </span>completions = [<span class="hljs-string">" blue."</span>, <span class="hljs-string">" in the sky."</span>] <span class="hljs-comment"># not used in the reward function, but the trainer will pass it</span> | |
| <span class="hljs-meta">>>> </span>completions_ids = [[<span class="hljs-number">6303</span>, <span class="hljs-number">13</span>], [<span class="hljs-number">304</span>, <span class="hljs-number">279</span>, <span class="hljs-number">12884</span>, <span class="hljs-number">13</span>]] | |
| <span class="hljs-meta">>>> </span>reward_func(prompts=prompts, completions=completions, completions_ids=completions_ids) | |
| [<span class="hljs-number">2.0</span>, <span class="hljs-number">4.0</span>]<!-- HTML_TAG_END --></pre></div> <h4 class="relative group"><a id="example-11-reward-longer-completions-based-on-the-number-of-characters" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#example-11-reward-longer-completions-based-on-the-number-of-characters"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Example 1.1: Reward longer completions (based on the number of characters)</span></h4> <p data-svelte-h="svelte-mgdbdy">Same as the previous example, but this time the reward function is based on the number of characters instead of tokens.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">def</span> <span class="hljs-title function_">reward_func</span>(<span class="hljs-params">completions, **kwargs</span>): | |
| <span class="hljs-string">"""Reward function that assigns higher scores to longer completions (in terms of character count)."""</span> | |
| <span class="hljs-keyword">return</span> [<span class="hljs-built_in">float</span>(<span class="hljs-built_in">len</span>(completion)) <span class="hljs-keyword">for</span> completion <span class="hljs-keyword">in</span> completions]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-17jijhc">You can test it as follows:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>prompts = [<span class="hljs-string">"The sky is"</span>, <span class="hljs-string">"The sun is"</span>] | |
| <span class="hljs-meta">>>> </span>completions = [<span class="hljs-string">" blue."</span>, <span class="hljs-string">" in the sky."</span>] | |
| <span class="hljs-meta">>>> </span>completions_ids = [[<span class="hljs-number">6303</span>, <span class="hljs-number">13</span>], [<span class="hljs-number">304</span>, <span class="hljs-number">279</span>, <span class="hljs-number">12884</span>, <span class="hljs-number">13</span>]] <span class="hljs-comment"># not used in the reward function, but the trainer will pass it</span> | |
| <span class="hljs-meta">>>> </span>reward_func(prompts=prompts, completions=completions, completions_ids=completions_ids) | |
| [<span class="hljs-number">6.0</span>, <span class="hljs-number">12.0</span>]<!-- HTML_TAG_END --></pre></div> <h4 class="relative group"><a id="example-2-reward-completions-with-a-specific-format" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#example-2-reward-completions-with-a-specific-format"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Example 2: Reward completions with a specific format</span></h4> <p data-svelte-h="svelte-1ublx2j">Below is an example of a reward function that checks if the completion has a specific format. This example is inspired by the <em>format reward</em> function used in the paper <a href="https://huggingface.co/papers/2501.12948" rel="nofollow">DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning</a>. | |
| It is designed for a conversational format, where prompts and completions consist of structured messages.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">import</span> re | |
| <span class="hljs-keyword">def</span> <span class="hljs-title function_">format_reward_func</span>(<span class="hljs-params">completions, **kwargs</span>): | |
| <span class="hljs-string">"""Reward function that checks if the completion has a specific format."""</span> | |
| pattern = <span class="hljs-string">r"^<think>.*?</think><answer>.*?</answer>$"</span> | |
| completion_contents = [completion[<span class="hljs-number">0</span>][<span class="hljs-string">"content"</span>] <span class="hljs-keyword">for</span> completion <span class="hljs-keyword">in</span> completions] | |
| matches = [re.<span class="hljs-keyword">match</span>(pattern, content) <span class="hljs-keyword">for</span> content <span class="hljs-keyword">in</span> completion_contents] | |
| <span class="hljs-keyword">return</span> [<span class="hljs-number">1.0</span> <span class="hljs-keyword">if</span> <span class="hljs-keyword">match</span> <span class="hljs-keyword">else</span> <span class="hljs-number">0.0</span> <span class="hljs-keyword">for</span> <span class="hljs-keyword">match</span> <span class="hljs-keyword">in</span> matches]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-19axqz5">You can test this function as follows:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>prompts = [ | |
| <span class="hljs-meta">... </span> [{<span class="hljs-string">"role"</span>: <span class="hljs-string">"assistant"</span>, <span class="hljs-string">"content"</span>: <span class="hljs-string">"What is the result of (1 + 2) * 4?"</span>}], | |
| <span class="hljs-meta">... </span> [{<span class="hljs-string">"role"</span>: <span class="hljs-string">"assistant"</span>, <span class="hljs-string">"content"</span>: <span class="hljs-string">"What is the result of (3 + 1) * 2?"</span>}], | |
| <span class="hljs-meta">... </span>] | |
| <span class="hljs-meta">>>> </span>completions = [ | |
| <span class="hljs-meta">... </span> [{<span class="hljs-string">"role"</span>: <span class="hljs-string">"assistant"</span>, <span class="hljs-string">"content"</span>: <span class="hljs-string">"<think>The sum of 1 and 2 is 3, which we multiply by 4 to get 12.</think><answer>(1 + 2) * 4 = 12</answer>"</span>}], | |
| <span class="hljs-meta">... </span> [{<span class="hljs-string">"role"</span>: <span class="hljs-string">"assistant"</span>, <span class="hljs-string">"content"</span>: <span class="hljs-string">"The sum of 3 and 1 is 4, which we multiply by 2 to get 8. So (3 + 1) * 2 = 8."</span>}], | |
| <span class="hljs-meta">... </span>] | |
| <span class="hljs-meta">>>> </span>format_reward_func(prompts=prompts, completions=completions) | |
| [<span class="hljs-number">1.0</span>, <span class="hljs-number">0.0</span>]<!-- HTML_TAG_END --></pre></div> <h4 class="relative group"><a id="example-3-reward-completions-based-on-a-reference" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#example-3-reward-completions-based-on-a-reference"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Example 3: Reward completions based on a reference</span></h4> <p data-svelte-h="svelte-d6x9c8">Below is an example of a reward function that checks if the completion is correct. This example is inspired by the <em>accuracy reward</em> function used in the paper <a href="https://huggingface.co/papers/2501.12948" rel="nofollow">DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning</a>. | |
| This example is designed for <a href="dataset_formats#standard">standard format</a>, where the dataset contains a column named <code>ground_truth</code>.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">import</span> re | |
| <span class="hljs-keyword">def</span> <span class="hljs-title function_">reward_func</span>(<span class="hljs-params">completions, ground_truth, **kwargs</span>): | |
| <span class="hljs-comment"># Regular expression to capture content inside \boxed{}</span> | |
| matches = [re.search(<span class="hljs-string">r"\\boxed\{(.*?)\}"</span>, completion) <span class="hljs-keyword">for</span> completion <span class="hljs-keyword">in</span> completions] | |
| contents = [<span class="hljs-keyword">match</span>.group(<span class="hljs-number">1</span>) <span class="hljs-keyword">if</span> <span class="hljs-keyword">match</span> <span class="hljs-keyword">else</span> <span class="hljs-string">""</span> <span class="hljs-keyword">for</span> <span class="hljs-keyword">match</span> <span class="hljs-keyword">in</span> matches] | |
| <span class="hljs-comment"># Reward 1 if the content is the same as the ground truth, 0 otherwise</span> | |
| <span class="hljs-keyword">return</span> [<span class="hljs-number">1.0</span> <span class="hljs-keyword">if</span> c == gt <span class="hljs-keyword">else</span> <span class="hljs-number">0.0</span> <span class="hljs-keyword">for</span> c, gt <span class="hljs-keyword">in</span> <span class="hljs-built_in">zip</span>(contents, ground_truth)]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-19axqz5">You can test this function as follows:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-meta">>>> </span>prompts = [<span class="hljs-string">"Problem: Solve the equation $2x + 3 = 7$. Solution:"</span>, <span class="hljs-string">"Problem: Solve the equation $3x - 5 = 10$."</span>] | |
| <span class="hljs-meta">>>> </span>completions = [<span class="hljs-string">r" The solution is \boxed{2}."</span>, <span class="hljs-string">r" The solution is \boxed{6}."</span>] | |
| <span class="hljs-meta">>>> </span>ground_truth = [<span class="hljs-string">"2"</span>, <span class="hljs-string">"5"</span>] | |
| <span class="hljs-meta">>>> </span>reward_func(prompts=prompts, completions=completions, ground_truth=ground_truth) | |
| [<span class="hljs-number">1.0</span>, <span class="hljs-number">0.0</span>]<!-- HTML_TAG_END --></pre></div> <h4 class="relative group"><a id="example-4-multi-task-reward-functions" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#example-4-multi-task-reward-functions"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Example 4: Multi-task reward functions</span></h4> <p data-svelte-h="svelte-11mflyk">Below is an example of using multiple reward functions in the <a href="/docs/trl/pr_4331/en/grpo_trainer#trl.GRPOTrainer">GRPOTrainer</a>. In this example, we define two task-specific reward functions: <code>math_reward_func</code> and <code>coding_reward_func</code>. The <code>math_reward_func</code> rewards math problems based on their correctness, while the <code>coding_reward_func</code> rewards coding problems based on whether the solution works.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> datasets <span class="hljs-keyword">import</span> Dataset | |
| <span class="hljs-keyword">from</span> trl <span class="hljs-keyword">import</span> GRPOTrainer | |
| <span class="hljs-comment"># Define a dataset that contains both math and coding problems</span> | |
| dataset = Dataset.from_list( | |
| [ | |
| {<span class="hljs-string">"prompt"</span>: <span class="hljs-string">"What is 2+2?"</span>, <span class="hljs-string">"task"</span>: <span class="hljs-string">"math"</span>}, | |
| {<span class="hljs-string">"prompt"</span>: <span class="hljs-string">"Write a function that returns the sum of two numbers."</span>, <span class="hljs-string">"task"</span>: <span class="hljs-string">"code"</span>}, | |
| {<span class="hljs-string">"prompt"</span>: <span class="hljs-string">"What is 3*4?"</span>, <span class="hljs-string">"task"</span>: <span class="hljs-string">"math"</span>}, | |
| {<span class="hljs-string">"prompt"</span>: <span class="hljs-string">"Write a function that returns the product of two numbers."</span>, <span class="hljs-string">"task"</span>: <span class="hljs-string">"code"</span>}, | |
| ] | |
| ) | |
| <span class="hljs-comment"># Math-specific reward function</span> | |
| <span class="hljs-keyword">def</span> <span class="hljs-title function_">math_reward_func</span>(<span class="hljs-params">prompts, completions, task, **kwargs</span>): | |
| rewards = [] | |
| <span class="hljs-keyword">for</span> prompt, completion, t <span class="hljs-keyword">in</span> <span class="hljs-built_in">zip</span>(prompts, completions, task): | |
| <span class="hljs-keyword">if</span> t == <span class="hljs-string">"math"</span>: | |
| <span class="hljs-comment"># Calculate math-specific reward</span> | |
| correct = check_math_solution(prompt, completion) | |
| reward = <span class="hljs-number">1.0</span> <span class="hljs-keyword">if</span> correct <span class="hljs-keyword">else</span> -<span class="hljs-number">1.0</span> | |
| rewards.append(reward) | |
| <span class="hljs-keyword">else</span>: | |
| <span class="hljs-comment"># Return None for non-math tasks</span> | |
| rewards.append(<span class="hljs-literal">None</span>) | |
| <span class="hljs-keyword">return</span> rewards | |
| <span class="hljs-comment"># Coding-specific reward function</span> | |
| <span class="hljs-keyword">def</span> <span class="hljs-title function_">coding_reward_func</span>(<span class="hljs-params">prompts, completions, task, **kwargs</span>): | |
| rewards = [] | |
| <span class="hljs-keyword">for</span> prompt, completion, t <span class="hljs-keyword">in</span> <span class="hljs-built_in">zip</span>(prompts, completions, task): | |
| <span class="hljs-keyword">if</span> t == <span class="hljs-string">"coding"</span>: | |
| <span class="hljs-comment"># Calculate coding-specific reward</span> | |
| works = test_code_solution(prompt, completion) | |
| reward = <span class="hljs-number">1.0</span> <span class="hljs-keyword">if</span> works <span class="hljs-keyword">else</span> -<span class="hljs-number">1.0</span> | |
| rewards.append(reward) | |
| <span class="hljs-keyword">else</span>: | |
| <span class="hljs-comment"># Return None for non-coding tasks</span> | |
| rewards.append(<span class="hljs-literal">None</span>) | |
| <span class="hljs-keyword">return</span> rewards | |
| <span class="hljs-comment"># Use both task-specific reward functions</span> | |
| trainer = GRPOTrainer( | |
| model=<span class="hljs-string">"Qwen/Qwen2-0.5B-Instruct"</span>, | |
| reward_funcs=[math_reward_func, coding_reward_func], | |
| train_dataset=dataset, | |
| ) | |
| trainer.train()<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1c61wht">In this example, the <code>math_reward_func</code> and <code>coding_reward_func</code> are designed to work with a mixed dataset that contains both math and coding problems. The <code>task</code> column in the dataset is used to determine which reward function to apply to each problem. If there is no relevant reward function for a sample in the dataset, the reward function will return <code>None</code>, and the <a href="/docs/trl/pr_4331/en/grpo_trainer#trl.GRPOTrainer">GRPOTrainer</a> will continue with the valid functions and tasks. This allows the <a href="/docs/trl/pr_4331/en/grpo_trainer#trl.GRPOTrainer">GRPOTrainer</a> to handle multiple reward functions with different applicability.</p> <p data-svelte-h="svelte-usdsiy">Note that the <a href="/docs/trl/pr_4331/en/grpo_trainer#trl.GRPOTrainer">GRPOTrainer</a> will ignore the <code>None</code> rewards returned by the reward functions and only consider the rewards returned by the relevant functions. This ensures that the model is trained on the relevant tasks and ignores the tasks for which there is no relevant reward function.</p> <h4 class="relative group"><a id="passing-the-reward-function-to-the-trainer" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#passing-the-reward-function-to-the-trainer"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Passing the reward function to the trainer</span></h4> <p data-svelte-h="svelte-ya9ro8">To use your custom reward function, pass it to the <a href="/docs/trl/pr_4331/en/grpo_trainer#trl.GRPOTrainer">GRPOTrainer</a> as follows:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> trl <span class="hljs-keyword">import</span> GRPOTrainer | |
| trainer = GRPOTrainer( | |
| reward_funcs=reward_func, | |
| ..., | |
| )<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1nd0tpp">If you have multiple reward functions, you can pass them as a list:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> trl <span class="hljs-keyword">import</span> GRPOTrainer | |
| trainer = GRPOTrainer( | |
| reward_funcs=[reward_func1, reward_func2], | |
| ..., | |
| )<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-104ub32">and the reward will be computed as the sum of the rewards from each function, or the weighted sum if <code>reward_weights</code> is provided in the config.</p> <p data-svelte-h="svelte-1ig5e5p">Note that <a href="/docs/trl/pr_4331/en/grpo_trainer#trl.GRPOTrainer">GRPOTrainer</a> supports multiple reward functions of different types. See the parameters documentation for more details.</p> <h2 class="relative group"><a id="vision-language-model-vlm-training" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#vision-language-model-vlm-training"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Vision-Language Model (VLM) Training</span></h2> <p data-svelte-h="svelte-1atcqx7">GRPO supports training Vision-Language Models (VLMs) on multimodal datasets containing both text and images.</p> <h3 class="relative group"><a id="supported-models" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#supported-models"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Supported Models</span></h3> <p data-svelte-h="svelte-1clwarl">Tested with:</p> <ul data-svelte-h="svelte-1dxn67f"><li><strong>Gemma3</strong> — e.g., <code>google/gemma-3-4b-it</code></li> <li><strong>LLaVA-NeXT</strong> — e.g., <code>llava-hf/llava-v1.6-mistral-7b-hf</code></li> <li><strong>Qwen2-VL</strong> — e.g., <code>Qwen/Qwen2-VL-2B-Instruct</code></li> <li><strong>Qwen2.5-VL</strong> — e.g., <code>Qwen/Qwen2.5-VL-3B-Instruct</code></li> <li><strong>SmolVLM2</strong> — e.g., <code>HuggingFaceTB/SmolVLM2-2.2B-Instruct</code></li></ul> <blockquote class="tip" data-svelte-h="svelte-pwu269"><p>Compatibility with all VLMs is not guaranteed. If you believe a model should be supported, feel free to open an issue on GitHub — or better yet, submit a pull request with the required changes.</p></blockquote> <h3 class="relative group"><a id="quick-start" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#quick-start"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Quick Start</span></h3> <p data-svelte-h="svelte-1e86bb2">Use <a href="https://github.com/huggingface/trl/blob/main/examples/scripts/grpo_vlm.py" rel="nofollow">grpo_vlm.py</a> to fine-tune a VLM. Example command for training on <a href="https://huggingface.co/datasets/lmms-lab/multimodal-open-r1-8k-verified" rel="nofollow"><code>lmms-lab/multimodal-open-r1-8k-verified</code></a>:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->accelerate launch \ | |
| --config_file=examples/accelerate_configs/deepspeed_zero3.yaml \ | |
| examples/scripts/grpo_vlm.py \ | |
| --model_name_or_path Qwen/Qwen2.5-VL-3B-Instruct \ | |
| --output_dir grpo-Qwen2.5-VL-3B-Instruct \ | |
| --learning_rate 1e-5 \ | |
| --gradient_checkpointing \ | |
| --dtype bfloat16 \ | |
| --max_prompt_length 2048 \ | |
| --max_completion_length 1024 \ | |
| --use_vllm \ | |
| --vllm_mode colocate \ | |
| --use_peft \ | |
| --lora_target_modules <span class="hljs-string">"q_proj"</span>, <span class="hljs-string">"v_proj"</span> \ | |
| --log_completions<!-- HTML_TAG_END --></pre></div> <h3 class="relative group"><a id="configuration-tips" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#configuration-tips"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Configuration Tips</span></h3> <blockquote class="tip"><p data-svelte-h="svelte-1up3tkl">For VLMs, truncating may remove image tokens, leading to errors during training. To avoid this, set <code>max_prompt_length=None</code> in the <a href="/docs/trl/pr_4331/en/grpo_trainer#trl.GRPOConfig">GRPOConfig</a>. This allows the model to process the full sequence length without truncating image tokens.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->GRPOConfig(max_prompt_length=<span class="hljs-literal">None</span>, ...)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1kut4fm">Only use <code>max_prompt_length</code> when you’ve verified that truncation won’t remove image tokens for the entire dataset.</p></blockquote> <ul data-svelte-h="svelte-kv9quh"><li>Use LoRA on vision-language projection layers</li> <li>Enable 4-bit quantization to reduce memory usage</li> <li>VLMs are memory-intensive — start with smaller batch sizes</li> <li>Most models are compatible with vLLM (<code>server</code> and <code>colocate</code> modes)</li></ul> <h3 class="relative group"><a id="dataset-format" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#dataset-format"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Dataset Format</span></h3> <p data-svelte-h="svelte-985b2u">Each training sample should include:</p> <ul data-svelte-h="svelte-1kosedy"><li><code>prompt</code>: Text formatted via the processor’s chat template</li> <li><code>image</code>/<code>images</code>: PIL Image or list of PIL Images</li></ul> <p data-svelte-h="svelte-t3zbfe">The trainer automatically handles image-to-tensor conversion via the model’s image processor.</p> <h2 class="relative group"><a id="trl.GRPOTrainer" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOTrainer"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>GRPOTrainer</span></h2> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"> <div><span class="group flex space-x-1.5 items-center text-gray-800 bg-gradient-to-r rounded-tr-lg -mt-4 -ml-4 pt-3 px-2.5" id="trl.GRPOTrainer"><!-- HTML_TAG_START --><h3 class="!m-0"><span class="flex-1 break-all md:text-lg bg-gradient-to-r px-2.5 py-1.5 rounded-xl from-indigo-50/70 to-white dark:from-gray-900 dark:to-gray-950 dark:text-indigo-300 text-indigo-700"><svg class="mr-1.5 text-indigo-500 inline-block -mt-0.5" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" focusable="false" role="img" width=".8em" height=".8em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 24 24"><path class="uim-quaternary" d="M20.23 7.24L12 12L3.77 7.24a1.98 1.98 0 0 1 .7-.71L11 2.76c.62-.35 1.38-.35 2 0l6.53 3.77c.29.173.531.418.7.71z" opacity=".25" fill="currentColor"></path><path class="uim-tertiary" d="M12 12v9.5a2.09 2.09 0 0 1-.91-.21L4.5 17.48a2.003 2.003 0 0 1-1-1.73v-7.5a2.06 2.06 0 0 1 .27-1.01L12 12z" opacity=".5" fill="currentColor"></path><path class="uim-primary" d="M20.5 8.25v7.5a2.003 2.003 0 0 1-1 1.73l-6.62 3.82c-.275.13-.576.198-.88.2V12l8.23-4.76c.175.308.268.656.27 1.01z" fill="currentColor"></path></svg><span class="font-light">class</span> <span class="font-medium">trl.</span><span class="font-semibold">GRPOTrainer</span></span></h3><!-- HTML_TAG_END --> <a id="trl.GRPOTrainer" class="header-link invisible with-hover:group-hover:visible pr-2" href="#trl.GRPOTrainer"><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></a> <a class="!ml-auto !text-gray-400 !no-underline text-sm flex items-center" href="https://github.com/huggingface/trl/blob/vr_4331/trl/trainer/grpo_trainer.py#L115" target="_blank"><span data-svelte-h="svelte-1kd6by1"><</span> <span class="hidden md:block mx-0.5 hover:!underline" data-svelte-h="svelte-122apf4">source</span> <span data-svelte-h="svelte-x0xyl0">></span></a></span> <p class="font-mono text-xs md:text-sm !leading-relaxed !my-6"><span data-svelte-h="svelte-8mvn6a">(</span> <span class="comma cursor-pointer"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">model<span class="opacity-60">: str | transformers.modeling_utils.PreTrainedModel</span></span> </span><span class="comma cursor-pointer"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">reward_funcs<span class="opacity-60">: str | transformers.modeling_utils.PreTrainedModel | collections.abc.Callable[[list, list], list[float]] | list[str | transformers.modeling_utils.PreTrainedModel | collections.abc.Callable[[list, list], list[float]]]</span></span> </span><span class="comma cursor-pointer"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">args<span class="opacity-60">: trl.trainer.grpo_config.GRPOConfig | None = None</span></span> </span><span class="comma cursor-pointer"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">train_dataset<span class="opacity-60">: datasets.arrow_dataset.Dataset | datasets.iterable_dataset.IterableDataset | None = None</span></span> </span><span class="comma cursor-pointer"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">eval_dataset<span class="opacity-60">: datasets.arrow_dataset.Dataset | datasets.iterable_dataset.IterableDataset | dict[str, datasets.arrow_dataset.Dataset | datasets.iterable_dataset.IterableDataset] | None = None</span></span> </span><span class="comma cursor-pointer"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">processing_class<span class="opacity-60">: transformers.tokenization_utils_base.PreTrainedTokenizerBase | transformers.processing_utils.ProcessorMixin | None = None</span></span> </span><span class="comma cursor-pointer"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">reward_processing_classes<span class="opacity-60">: transformers.tokenization_utils_base.PreTrainedTokenizerBase | list[transformers.tokenization_utils_base.PreTrainedTokenizerBase] | None = None</span></span> </span><span class="comma cursor-pointer"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">callbacks<span class="opacity-60">: list[transformers.trainer_callback.TrainerCallback] | None = None</span></span> </span><span class="comma cursor-pointer"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">optimizers<span class="opacity-60">: tuple = (None, None)</span></span> </span><span class="comma cursor-pointer"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">peft_config<span class="opacity-60">: PeftConfig | None = None</span></span> </span><span class="comma cursor-pointer"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">rollout_func<span class="opacity-60">: collections.abc.Callable[[list[str], typing.Any, typing.Any], dict[str, typing.Any]] | None = None</span></span> </span> <span data-svelte-h="svelte-1jq0pl7">)</span> </p> <div class="!mb-10 relative docstring-details "> <p class="flex items-center font-semibold !mt-2 !mb-2 text-gray-800" data-svelte-h="svelte-lt6pb6">Parameters <span class="flex-auto border-t-2 border-gray-100 dark:border-gray-700 ml-3"></span></p> <ul class="px-2"><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOTrainer.model" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOTrainer.model"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>model</strong> (<code>str | PreTrainedModel</code>) — | |
| Model to be trained. Can be either:</p> | |
| <ul> | |
| <li>A string, being the <em>model id</em> of a pretrained model hosted inside a model repo on huggingface.co, or a | |
| path to a <em>directory</em> containing model weights saved using | |
| <a href="https://huggingface.co/docs/transformers/main/en/main_classes/model#transformers.PreTrainedModel.save_pretrained" rel="nofollow">save_pretrained</a>, e.g., <code>'./my_model_directory/'</code>. The model is loaded | |
| using <a href="https://huggingface.co/docs/transformers/main/en/model_doc/auto#transformers.AutoModelForCausalLM.from_pretrained" rel="nofollow">from_pretrained</a> with the keyword arguments in | |
| <code>args.model_init_kwargs</code>.</li> | |
| <li>A <a href="https://huggingface.co/docs/transformers/main/en/main_classes/model#transformers.PreTrainedModel" rel="nofollow">PreTrainedModel</a> object. Only causal language models are supported.</li> | |
| </ul><!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOTrainer.reward_funcs" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOTrainer.reward_funcs"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>reward_funcs</strong> (<code>RewardFunc | list[RewardFunc]</code>) — | |
| Reward functions to be used for computing the rewards. To compute the rewards, we call all the reward | |
| functions with the prompts and completions and sum the rewards. Can be either:</p> | |
| <ul> | |
| <li> | |
| <p>A single reward function, such as:</p> | |
| <ul> | |
| <li> | |
| <p>A string: The <em>model ID</em> of a pretrained model hosted inside a model repo on huggingface.co, or a | |
| path to a <em>directory</em> containing model weights saved using | |
| <a href="https://huggingface.co/docs/transformers/main/en/main_classes/model#transformers.PreTrainedModel.save_pretrained" rel="nofollow">save_pretrained</a>, e.g., <code>'./my_model_directory/'</code>. The model is loaded | |
| using <a href="https://huggingface.co/docs/transformers/main/en/model_doc/auto#transformers.AutoModelForSequenceClassification.from_pretrained" rel="nofollow">from_pretrained</a> with <code>num_labels=1</code> and the | |
| keyword arguments in <code>args.model_init_kwargs</code>.</p> | |
| </li> | |
| <li> | |
| <p>A <a href="https://huggingface.co/docs/transformers/main/en/main_classes/model#transformers.PreTrainedModel" rel="nofollow">PreTrainedModel</a> object: Only sequence classification models are supported.</p> | |
| </li> | |
| <li> | |
| <p>A custom reward function: The function is provided with the prompts and the generated completions, | |
| plus any additional columns in the dataset. It should return a list of rewards. Custom reward | |
| functions can also return <code>None</code> when the reward is not applicable to those samples. This is useful | |
| for multi-task training where different reward functions apply to different types of samples. When a | |
| reward function returns <code>None</code> for a sample, that reward function is excluded from the reward | |
| calculation for that sample. For more details, see <a href="#using-a-custom-reward-function">Using a custom reward | |
| function</a>.</p> | |
| <p>The trainer’s state is also passed to the reward function. The trainer’s state is an instance of | |
| <a href="https://huggingface.co/docs/transformers/main/en/main_classes/callback#transformers.TrainerState" rel="nofollow">TrainerState</a> and can be accessed by accessing the <code>trainer_state</code> argument to the | |
| reward function’s signature.</p> | |
| </li> | |
| </ul> | |
| </li> | |
| <li> | |
| <p>A list of reward functions, where each item can independently be any of the above types. Mixing different | |
| types within the list (e.g., a string model ID and a custom reward function) is allowed.</p> | |
| </li> | |
| </ul><!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOTrainer.args" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOTrainer.args"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>args</strong> (<a href="/docs/trl/pr_4331/en/grpo_trainer#trl.GRPOConfig">GRPOConfig</a>, <em>optional</em>) — | |
| Configuration for this trainer. If <code>None</code>, a default configuration is used.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOTrainer.train_dataset" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOTrainer.train_dataset"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>train_dataset</strong> (<a href="https://huggingface.co/docs/datasets/main/en/package_reference/main_classes#datasets.Dataset" rel="nofollow">Dataset</a> or <a href="https://huggingface.co/docs/datasets/main/en/package_reference/main_classes#datasets.IterableDataset" rel="nofollow">IterableDataset</a>) — | |
| Dataset to use for training. It must include a column <code>"prompt"</code>. Any additional columns in the dataset is | |
| ignored. The format of the samples can be either:</p> | |
| <ul> | |
| <li><a href="dataset_formats#standard">Standard</a>: Each sample contains plain text.</li> | |
| <li><a href="dataset_formats#conversational">Conversational</a>: Each sample contains structured messages (e.g., role | |
| and content).</li> | |
| </ul><!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOTrainer.eval_dataset" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOTrainer.eval_dataset"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>eval_dataset</strong> (<a href="https://huggingface.co/docs/datasets/main/en/package_reference/main_classes#datasets.Dataset" rel="nofollow">Dataset</a>, <a href="https://huggingface.co/docs/datasets/main/en/package_reference/main_classes#datasets.IterableDataset" rel="nofollow">IterableDataset</a> or <code>dict[str, Dataset | IterableDataset]</code>) — | |
| Dataset to use for evaluation. It must meet the same requirements as <code>train_dataset</code>.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOTrainer.processing_class" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOTrainer.processing_class"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>processing_class</strong> (<a href="https://huggingface.co/docs/transformers/main/en/internal/tokenization_utils#transformers.PreTrainedTokenizerBase" rel="nofollow">PreTrainedTokenizerBase</a>, <a href="https://huggingface.co/docs/transformers/main/en/main_classes/processors#transformers.ProcessorMixin" rel="nofollow">ProcessorMixin</a>, <em>optional</em>) — | |
| Processing class used to process the data. The padding side must be set to “left”. If <code>None</code>, the | |
| processing class is loaded from the model’s name with <a href="https://huggingface.co/docs/transformers/main/en/model_doc/auto#transformers.AutoProcessor.from_pretrained" rel="nofollow">from_pretrained</a>. A | |
| padding token, <code>tokenizer.pad_token</code>, must be set. If the processing class has not set a padding token, | |
| <code>tokenizer.eos_token</code> will be used as the default.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOTrainer.reward_processing_classes" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOTrainer.reward_processing_classes"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>reward_processing_classes</strong> (<a href="https://huggingface.co/docs/transformers/main/en/internal/tokenization_utils#transformers.PreTrainedTokenizerBase" rel="nofollow">PreTrainedTokenizerBase</a> or <code>list[PreTrainedTokenizerBase]</code>, <em>optional</em>) — | |
| Processing classes corresponding to the reward functions specified in <code>reward_funcs</code>. Can be either:</p> | |
| <ul> | |
| <li>A single processing class: Used when <code>reward_funcs</code> contains only one reward function.</li> | |
| <li>A list of processing classes: Must match the order and length of the reward functions in <code>reward_funcs</code>. | |
| If set to <code>None</code>, or if an element of the list corresponding to a <a href="https://huggingface.co/docs/transformers/main/en/main_classes/model#transformers.PreTrainedModel" rel="nofollow">PreTrainedModel</a> is | |
| <code>None</code>, the tokenizer for the model is automatically loaded using | |
| <a href="https://huggingface.co/docs/transformers/main/en/model_doc/auto#transformers.AutoTokenizer.from_pretrained" rel="nofollow">from_pretrained</a>. For elements in <code>reward_funcs</code> that are custom reward | |
| functions (not <a href="https://huggingface.co/docs/transformers/main/en/main_classes/model#transformers.PreTrainedModel" rel="nofollow">PreTrainedModel</a>), the corresponding entries in <code>reward_processing_classes</code> | |
| are ignored.</li> | |
| </ul><!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOTrainer.callbacks" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOTrainer.callbacks"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>callbacks</strong> (list of <a href="https://huggingface.co/docs/transformers/main/en/main_classes/callback#transformers.TrainerCallback" rel="nofollow">TrainerCallback</a>, <em>optional</em>) — | |
| List of callbacks to customize the training loop. Will add those to the list of default callbacks detailed | |
| in <a href="https://huggingface.co/docs/transformers/main_classes/callback" rel="nofollow">here</a>.</p> | |
| <p>If you want to remove one of the default callbacks used, use the <a href="https://huggingface.co/docs/transformers/main/en/main_classes/trainer#transformers.Trainer.remove_callback" rel="nofollow">remove_callback</a> | |
| method.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOTrainer.optimizers" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOTrainer.optimizers"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>optimizers</strong> (<code>tuple[torch.optim.Optimizer, torch.optim.lr_scheduler.LambdaLR]</code>, <em>optional</em>, defaults to <code>(None, None)</code>) — | |
| A tuple containing the optimizer and the scheduler to use. Will default to an instance of <code>AdamW</code> on your | |
| model and a scheduler given by <code>get_linear_schedule_with_warmup</code> controlled by <code>args</code>.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOTrainer.peft_config" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOTrainer.peft_config"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>peft_config</strong> (<a href="https://huggingface.co/docs/peft/main/en/package_reference/config#peft.PeftConfig" rel="nofollow">PeftConfig</a>, <em>optional</em>) — | |
| PEFT configuration used to wrap the model. If <code>None</code>, the model is not wrapped.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOTrainer.rollout_func" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOTrainer.rollout_func"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>rollout_func</strong> (<code>RolloutFunc</code>, <em>optional</em>) — | |
| Function to use for generating completions. It must take prompts, args, and processing_class as parameters | |
| and return a dict with <code>"prompt_ids"</code>, <code>"completion_ids"</code>, and <code>"logprobs"</code> fields. Any other fields that | |
| are forwarded to the reward functions. This feature is experimental and may change or be removed at any | |
| time without prior notice.<!-- HTML_TAG_END --> </span></span> </li></ul> </div></div> <p data-svelte-h="svelte-udysum">Trainer for the Group Relative Policy Optimization (GRPO) method. This algorithm was initially proposed in the | |
| paper <a href="https://huggingface.co/papers/2402.03300" rel="nofollow">DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language | |
| Models</a>.</p> <div class="relative group rounded-md"><a id="trl.GRPOTrainer.example" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOTrainer.example"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <p data-svelte-h="svelte-11lpom8">Example:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> datasets <span class="hljs-keyword">import</span> load_dataset | |
| <span class="hljs-keyword">from</span> trl <span class="hljs-keyword">import</span> GRPOTrainer | |
| dataset = load_dataset(<span class="hljs-string">"trl-lib/tldr"</span>, split=<span class="hljs-string">"train"</span>) | |
| <span class="hljs-keyword">def</span> <span class="hljs-title function_">reward_func</span>(<span class="hljs-params">completions, **kwargs</span>): | |
| <span class="hljs-comment"># Dummy reward function that rewards completions with more unique letters.</span> | |
| <span class="hljs-keyword">return</span> [<span class="hljs-built_in">float</span>(<span class="hljs-built_in">len</span>(<span class="hljs-built_in">set</span>(completion))) <span class="hljs-keyword">for</span> completion <span class="hljs-keyword">in</span> completions] | |
| trainer = GRPOTrainer( | |
| model=<span class="hljs-string">"Qwen/Qwen2-0.5B-Instruct"</span>, | |
| reward_funcs=reward_func, | |
| train_dataset=dataset, | |
| ) | |
| trainer.train()<!-- HTML_TAG_END --></pre></div></div> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"> <div><span class="group flex space-x-1.5 items-center text-gray-800 bg-gradient-to-r rounded-tr-lg -mt-4 -ml-4 pt-3 px-2.5" id="trl.GRPOTrainer.train"><!-- HTML_TAG_START --><h4 class="!m-0"><span class="flex-1 rounded-xl py-0.5 break-all bg-gradient-to-r from-blue-50/60 to-white dark:from-gray-900 dark:to-gray-950 text-blue-700 dark:text-blue-300 font-medium px-2"><svg width="1em" height="1em" viewBox="0 0 32 33" class="mr-1 inline-block -mt-0.5" xmlns="http://www.w3.org/2000/svg"><path d="M5.80566 18.3545C4.90766 17.4565 4.90766 16.0005 5.80566 15.1025L14.3768 6.53142C15.2748 5.63342 16.7307 5.63342 17.6287 6.53142L26.1999 15.1025C27.0979 16.0005 27.0979 17.4565 26.1999 18.3545L17.6287 26.9256C16.7307 27.8236 15.2748 27.8236 14.3768 26.9256L5.80566 18.3545Z" fill="currentColor" fill-opacity="0.25"/><path fill-rule="evenodd" clip-rule="evenodd" d="M16.4801 13.9619C16.4801 12.9761 16.7467 12.5436 16.9443 12.3296C17.1764 12.078 17.5731 11.8517 18.2275 11.707C18.8821 11.5623 19.638 11.5342 20.4038 11.5582C20.7804 11.57 21.1341 11.5932 21.4719 11.6156L21.5263 11.6193C21.8195 11.6389 22.1626 11.6618 22.4429 11.6618V7.40825C22.3209 7.40825 22.1219 7.39596 21.7544 7.37149C21.4202 7.34925 20.9976 7.32115 20.5371 7.30672C19.6286 7.27824 18.4672 7.29779 17.3093 7.55377C16.1512 7.8098 14.8404 8.33724 13.8181 9.4452C12.7612 10.5907 12.2266 12.1236 12.2266 13.9619V15.0127H10.6836V19.2662H12.2266V26.6332H16.4801V19.2662H20.3394V15.0127H16.4801V13.9619Z" fill="currentColor"/></svg>train</span></h4><!-- HTML_TAG_END --> <a id="trl.GRPOTrainer.train" class="header-link invisible with-hover:group-hover:visible pr-2" href="#trl.GRPOTrainer.train"><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></a> <a class="!ml-auto !text-gray-400 !no-underline text-sm flex items-center" href="https://github.com/huggingface/trl/blob/vr_4331/transformers/trainer.py#L2213" target="_blank"><span data-svelte-h="svelte-1kd6by1"><</span> <span class="hidden md:block mx-0.5 hover:!underline" data-svelte-h="svelte-122apf4">source</span> <span data-svelte-h="svelte-x0xyl0">></span></a></span> <p class="font-mono text-xs md:text-sm !leading-relaxed !my-6"><span data-svelte-h="svelte-8mvn6a">(</span> <span class="comma cursor-pointer"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">resume_from_checkpoint<span class="opacity-60">: typing.Union[str, bool, NoneType] = None</span></span> </span><span class="comma cursor-pointer"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">trial<span class="opacity-60">: typing.Union[ForwardRef('optuna.Trial'), dict[str, typing.Any], NoneType] = None</span></span> </span><span class="comma cursor-pointer"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">ignore_keys_for_eval<span class="opacity-60">: typing.Optional[list[str]] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">**kwargs<span class="opacity-60">: typing.Any</span></span> </span> <span data-svelte-h="svelte-1jq0pl7">)</span> </p> <div class="!mb-10 relative docstring-details "> <p class="flex items-center font-semibold !mt-2 !mb-2 text-gray-800" data-svelte-h="svelte-lt6pb6">Parameters <span class="flex-auto border-t-2 border-gray-100 dark:border-gray-700 ml-3"></span></p> <ul class="px-2"><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOTrainer.train.resume_from_checkpoint" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOTrainer.train.resume_from_checkpoint"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>resume_from_checkpoint</strong> (<code>str</code> or <code>bool</code>, <em>optional</em>) — | |
| If a <code>str</code>, local path to a saved checkpoint as saved by a previous instance of <code>Trainer</code>. If a | |
| <code>bool</code> and equals <code>True</code>, load the last checkpoint in <em>args.output_dir</em> as saved by a previous instance | |
| of <code>Trainer</code>. If present, training will resume from the model/optimizer/scheduler states loaded here.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOTrainer.train.trial" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOTrainer.train.trial"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>trial</strong> (<code>optuna.Trial</code> or <code>dict[str, Any]</code>, <em>optional</em>) — | |
| The trial run or the hyperparameter dictionary for hyperparameter search.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOTrainer.train.ignore_keys_for_eval" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOTrainer.train.ignore_keys_for_eval"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>ignore_keys_for_eval</strong> (<code>list[str]</code>, <em>optional</em>) — | |
| A list of keys in the output of your model (if it is a dictionary) that should be ignored when | |
| gathering predictions for evaluation during the training.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOTrainer.train.kwargs" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOTrainer.train.kwargs"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>kwargs</strong> (<code>dict[str, Any]</code>, <em>optional</em>) — | |
| Additional keyword arguments used to hide deprecated arguments<!-- HTML_TAG_END --> </span></span> </li></ul> </div></div> <p data-svelte-h="svelte-1cilnet">Main training entry point.</p></div> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"> <div><span class="group flex space-x-1.5 items-center text-gray-800 bg-gradient-to-r rounded-tr-lg -mt-4 -ml-4 pt-3 px-2.5" id="trl.GRPOTrainer.save_model"><!-- HTML_TAG_START --><h4 class="!m-0"><span class="flex-1 rounded-xl py-0.5 break-all bg-gradient-to-r from-blue-50/60 to-white dark:from-gray-900 dark:to-gray-950 text-blue-700 dark:text-blue-300 font-medium px-2"><svg width="1em" height="1em" viewBox="0 0 32 33" class="mr-1 inline-block -mt-0.5" xmlns="http://www.w3.org/2000/svg"><path d="M5.80566 18.3545C4.90766 17.4565 4.90766 16.0005 5.80566 15.1025L14.3768 6.53142C15.2748 5.63342 16.7307 5.63342 17.6287 6.53142L26.1999 15.1025C27.0979 16.0005 27.0979 17.4565 26.1999 18.3545L17.6287 26.9256C16.7307 27.8236 15.2748 27.8236 14.3768 26.9256L5.80566 18.3545Z" fill="currentColor" fill-opacity="0.25"/><path fill-rule="evenodd" clip-rule="evenodd" d="M16.4801 13.9619C16.4801 12.9761 16.7467 12.5436 16.9443 12.3296C17.1764 12.078 17.5731 11.8517 18.2275 11.707C18.8821 11.5623 19.638 11.5342 20.4038 11.5582C20.7804 11.57 21.1341 11.5932 21.4719 11.6156L21.5263 11.6193C21.8195 11.6389 22.1626 11.6618 22.4429 11.6618V7.40825C22.3209 7.40825 22.1219 7.39596 21.7544 7.37149C21.4202 7.34925 20.9976 7.32115 20.5371 7.30672C19.6286 7.27824 18.4672 7.29779 17.3093 7.55377C16.1512 7.8098 14.8404 8.33724 13.8181 9.4452C12.7612 10.5907 12.2266 12.1236 12.2266 13.9619V15.0127H10.6836V19.2662H12.2266V26.6332H16.4801V19.2662H20.3394V15.0127H16.4801V13.9619Z" fill="currentColor"/></svg>save_model</span></h4><!-- HTML_TAG_END --> <a id="trl.GRPOTrainer.save_model" class="header-link invisible with-hover:group-hover:visible pr-2" href="#trl.GRPOTrainer.save_model"><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></a> <a class="!ml-auto !text-gray-400 !no-underline text-sm flex items-center" href="https://github.com/huggingface/trl/blob/vr_4331/transformers/trainer.py#L4177" target="_blank"><span data-svelte-h="svelte-1kd6by1"><</span> <span class="hidden md:block mx-0.5 hover:!underline" data-svelte-h="svelte-122apf4">source</span> <span data-svelte-h="svelte-x0xyl0">></span></a></span> <p class="font-mono text-xs md:text-sm !leading-relaxed !my-6"><span data-svelte-h="svelte-8mvn6a">(</span> <span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">output_dir<span class="opacity-60">: typing.Optional[str] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">_internal_call<span class="opacity-60">: bool = False</span></span> </span> <span data-svelte-h="svelte-1jq0pl7">)</span> </p> <div class="!mb-10 relative docstring-details "> </div></div> <p data-svelte-h="svelte-r8h4ov">Will save the model, so you can reload it using <code>from_pretrained()</code>.</p> <p data-svelte-h="svelte-1e6bius">Will only save from the main process.</p></div> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"> <div><span class="group flex space-x-1.5 items-center text-gray-800 bg-gradient-to-r rounded-tr-lg -mt-4 -ml-4 pt-3 px-2.5" id="trl.GRPOTrainer.push_to_hub"><!-- HTML_TAG_START --><h4 class="!m-0"><span class="flex-1 rounded-xl py-0.5 break-all bg-gradient-to-r from-blue-50/60 to-white dark:from-gray-900 dark:to-gray-950 text-blue-700 dark:text-blue-300 font-medium px-2"><svg width="1em" height="1em" viewBox="0 0 32 33" class="mr-1 inline-block -mt-0.5" xmlns="http://www.w3.org/2000/svg"><path d="M5.80566 18.3545C4.90766 17.4565 4.90766 16.0005 5.80566 15.1025L14.3768 6.53142C15.2748 5.63342 16.7307 5.63342 17.6287 6.53142L26.1999 15.1025C27.0979 16.0005 27.0979 17.4565 26.1999 18.3545L17.6287 26.9256C16.7307 27.8236 15.2748 27.8236 14.3768 26.9256L5.80566 18.3545Z" fill="currentColor" fill-opacity="0.25"/><path fill-rule="evenodd" clip-rule="evenodd" d="M16.4801 13.9619C16.4801 12.9761 16.7467 12.5436 16.9443 12.3296C17.1764 12.078 17.5731 11.8517 18.2275 11.707C18.8821 11.5623 19.638 11.5342 20.4038 11.5582C20.7804 11.57 21.1341 11.5932 21.4719 11.6156L21.5263 11.6193C21.8195 11.6389 22.1626 11.6618 22.4429 11.6618V7.40825C22.3209 7.40825 22.1219 7.39596 21.7544 7.37149C21.4202 7.34925 20.9976 7.32115 20.5371 7.30672C19.6286 7.27824 18.4672 7.29779 17.3093 7.55377C16.1512 7.8098 14.8404 8.33724 13.8181 9.4452C12.7612 10.5907 12.2266 12.1236 12.2266 13.9619V15.0127H10.6836V19.2662H12.2266V26.6332H16.4801V19.2662H20.3394V15.0127H16.4801V13.9619Z" fill="currentColor"/></svg>push_to_hub</span></h4><!-- HTML_TAG_END --> <a id="trl.GRPOTrainer.push_to_hub" class="header-link invisible with-hover:group-hover:visible pr-2" href="#trl.GRPOTrainer.push_to_hub"><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></a> <a class="!ml-auto !text-gray-400 !no-underline text-sm flex items-center" href="https://github.com/huggingface/trl/blob/vr_4331/transformers/trainer.py#L5117" target="_blank"><span data-svelte-h="svelte-1kd6by1"><</span> <span class="hidden md:block mx-0.5 hover:!underline" data-svelte-h="svelte-122apf4">source</span> <span data-svelte-h="svelte-x0xyl0">></span></a></span> <p class="font-mono text-xs md:text-sm !leading-relaxed !my-6"><span data-svelte-h="svelte-8mvn6a">(</span> <span class="comma cursor-pointer"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">commit_message<span class="opacity-60">: typing.Optional[str] = 'End of training'</span></span> </span><span class="comma cursor-pointer"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">blocking<span class="opacity-60">: bool = True</span></span> </span><span class="comma cursor-pointer"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">token<span class="opacity-60">: typing.Optional[str] = None</span></span> </span><span class="comma cursor-pointer"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">revision<span class="opacity-60">: typing.Optional[str] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">**kwargs<span class="opacity-60"></span></span> </span> <span data-svelte-h="svelte-1jq0pl7">)</span> </p> <div class="!mb-10 relative docstring-details "> <p class="flex items-center font-semibold !mt-2 !mb-2 text-gray-800" data-svelte-h="svelte-lt6pb6">Parameters <span class="flex-auto border-t-2 border-gray-100 dark:border-gray-700 ml-3"></span></p> <ul class="px-2"><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOTrainer.push_to_hub.commit_message" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOTrainer.push_to_hub.commit_message"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>commit_message</strong> (<code>str</code>, <em>optional</em>, defaults to <code>"End of training"</code>) — | |
| Message to commit while pushing.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOTrainer.push_to_hub.blocking" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOTrainer.push_to_hub.blocking"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>blocking</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>True</code>) — | |
| Whether the function should return only when the <code>git push</code> has finished.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOTrainer.push_to_hub.token" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOTrainer.push_to_hub.token"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>token</strong> (<code>str</code>, <em>optional</em>, defaults to <code>None</code>) — | |
| Token with write permission to overwrite Trainer’s original args.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOTrainer.push_to_hub.revision" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOTrainer.push_to_hub.revision"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>revision</strong> (<code>str</code>, <em>optional</em>) — | |
| The git revision to commit from. Defaults to the head of the “main” branch.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOTrainer.push_to_hub.kwargs" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOTrainer.push_to_hub.kwargs"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>kwargs</strong> (<code>dict[str, Any]</code>, <em>optional</em>) — | |
| Additional keyword arguments passed along to <code>~Trainer.create_model_card</code>.<!-- HTML_TAG_END --> </span></span> </li></ul> </div></div> <p data-svelte-h="svelte-8tudwd">Upload <code>self.model</code> and <code>self.processing_class</code> to the 🤗 model hub on the repo <code>self.args.hub_model_id</code>.</p></div></div> <h2 class="relative group"><a id="trl.GRPOConfig" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>GRPOConfig</span></h2> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"> <div><span class="group flex space-x-1.5 items-center text-gray-800 bg-gradient-to-r rounded-tr-lg -mt-4 -ml-4 pt-3 px-2.5" id="trl.GRPOConfig"><!-- HTML_TAG_START --><h3 class="!m-0"><span class="flex-1 break-all md:text-lg bg-gradient-to-r px-2.5 py-1.5 rounded-xl from-indigo-50/70 to-white dark:from-gray-900 dark:to-gray-950 dark:text-indigo-300 text-indigo-700"><svg class="mr-1.5 text-indigo-500 inline-block -mt-0.5" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" focusable="false" role="img" width=".8em" height=".8em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 24 24"><path class="uim-quaternary" d="M20.23 7.24L12 12L3.77 7.24a1.98 1.98 0 0 1 .7-.71L11 2.76c.62-.35 1.38-.35 2 0l6.53 3.77c.29.173.531.418.7.71z" opacity=".25" fill="currentColor"></path><path class="uim-tertiary" d="M12 12v9.5a2.09 2.09 0 0 1-.91-.21L4.5 17.48a2.003 2.003 0 0 1-1-1.73v-7.5a2.06 2.06 0 0 1 .27-1.01L12 12z" opacity=".5" fill="currentColor"></path><path class="uim-primary" d="M20.5 8.25v7.5a2.003 2.003 0 0 1-1 1.73l-6.62 3.82c-.275.13-.576.198-.88.2V12l8.23-4.76c.175.308.268.656.27 1.01z" fill="currentColor"></path></svg><span class="font-light">class</span> <span class="font-medium">trl.</span><span class="font-semibold">GRPOConfig</span></span></h3><!-- HTML_TAG_END --> <a id="trl.GRPOConfig" class="header-link invisible with-hover:group-hover:visible pr-2" href="#trl.GRPOConfig"><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></a> <a class="!ml-auto !text-gray-400 !no-underline text-sm flex items-center" href="https://github.com/huggingface/trl/blob/vr_4331/trl/trainer/grpo_config.py#L22" target="_blank"><span data-svelte-h="svelte-1kd6by1"><</span> <span class="hidden md:block mx-0.5 hover:!underline" data-svelte-h="svelte-122apf4">source</span> <span data-svelte-h="svelte-x0xyl0">></span></a></span> <p class="font-mono text-xs md:text-sm !leading-relaxed !my-6"><span data-svelte-h="svelte-8mvn6a">(</span> <span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">output_dir<span class="opacity-60">: typing.Optional[str] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">overwrite_output_dir<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">do_train<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">do_eval<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">do_predict<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">eval_strategy<span class="opacity-60">: typing.Union[transformers.trainer_utils.IntervalStrategy, str] = 'no'</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">prediction_loss_only<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">per_device_train_batch_size<span class="opacity-60">: int = 8</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">per_device_eval_batch_size<span class="opacity-60">: int = 8</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">per_gpu_train_batch_size<span class="opacity-60">: typing.Optional[int] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">per_gpu_eval_batch_size<span class="opacity-60">: typing.Optional[int] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">gradient_accumulation_steps<span class="opacity-60">: int = 1</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">eval_accumulation_steps<span class="opacity-60">: typing.Optional[int] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">eval_delay<span class="opacity-60">: float = 0</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">torch_empty_cache_steps<span class="opacity-60">: typing.Optional[int] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">learning_rate<span class="opacity-60">: float = 1e-06</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">weight_decay<span class="opacity-60">: float = 0.0</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">adam_beta1<span class="opacity-60">: float = 0.9</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">adam_beta2<span class="opacity-60">: float = 0.999</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">adam_epsilon<span class="opacity-60">: float = 1e-08</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">max_grad_norm<span class="opacity-60">: float = 1.0</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">num_train_epochs<span class="opacity-60">: float = 3.0</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">max_steps<span class="opacity-60">: int = -1</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">lr_scheduler_type<span class="opacity-60">: typing.Union[transformers.trainer_utils.SchedulerType, str] = 'linear'</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">lr_scheduler_kwargs<span class="opacity-60">: typing.Union[dict[str, typing.Any], str] = <factory></span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">warmup_ratio<span class="opacity-60">: float = 0.0</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">warmup_steps<span class="opacity-60">: int = 0</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">log_level<span class="opacity-60">: str = 'passive'</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">log_level_replica<span class="opacity-60">: str = 'warning'</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">log_on_each_node<span class="opacity-60">: bool = True</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">logging_dir<span class="opacity-60">: typing.Optional[str] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">logging_strategy<span class="opacity-60">: typing.Union[transformers.trainer_utils.IntervalStrategy, str] = 'steps'</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">logging_first_step<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">logging_steps<span class="opacity-60">: float = 10</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">logging_nan_inf_filter<span class="opacity-60">: bool = True</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">save_strategy<span class="opacity-60">: typing.Union[transformers.trainer_utils.SaveStrategy, str] = 'steps'</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">save_steps<span class="opacity-60">: float = 500</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">save_total_limit<span class="opacity-60">: typing.Optional[int] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">save_safetensors<span class="opacity-60">: bool = True</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">save_on_each_node<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">save_only_model<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">restore_callback_states_from_checkpoint<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">no_cuda<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">use_cpu<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">use_mps_device<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">seed<span class="opacity-60">: int = 42</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">data_seed<span class="opacity-60">: typing.Optional[int] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">jit_mode_eval<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">bf16<span class="opacity-60">: bool | None = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">fp16<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">fp16_opt_level<span class="opacity-60">: str = 'O1'</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">half_precision_backend<span class="opacity-60">: str = 'auto'</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">bf16_full_eval<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">fp16_full_eval<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">tf32<span class="opacity-60">: typing.Optional[bool] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">local_rank<span class="opacity-60">: int = -1</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">ddp_backend<span class="opacity-60">: typing.Optional[str] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">tpu_num_cores<span class="opacity-60">: typing.Optional[int] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">tpu_metrics_debug<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">debug<span class="opacity-60">: typing.Union[str, list[transformers.debug_utils.DebugOption]] = ''</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">dataloader_drop_last<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">eval_steps<span class="opacity-60">: typing.Optional[float] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">dataloader_num_workers<span class="opacity-60">: int = 0</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">dataloader_prefetch_factor<span class="opacity-60">: typing.Optional[int] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">past_index<span class="opacity-60">: int = -1</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">run_name<span class="opacity-60">: typing.Optional[str] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">disable_tqdm<span class="opacity-60">: typing.Optional[bool] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">remove_unused_columns<span class="opacity-60">: bool | None = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">label_names<span class="opacity-60">: typing.Optional[list[str]] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">load_best_model_at_end<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">metric_for_best_model<span class="opacity-60">: typing.Optional[str] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">greater_is_better<span class="opacity-60">: typing.Optional[bool] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">ignore_data_skip<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">fsdp<span class="opacity-60">: typing.Union[list[transformers.trainer_utils.FSDPOption], str, NoneType] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">fsdp_min_num_params<span class="opacity-60">: int = 0</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">fsdp_config<span class="opacity-60">: typing.Union[dict[str, typing.Any], str, NoneType] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">fsdp_transformer_layer_cls_to_wrap<span class="opacity-60">: typing.Optional[str] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">accelerator_config<span class="opacity-60">: typing.Union[dict, str, NoneType] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">parallelism_config<span class="opacity-60">: typing.Optional[accelerate.parallelism_config.ParallelismConfig] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">deepspeed<span class="opacity-60">: typing.Union[dict, str, NoneType] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">label_smoothing_factor<span class="opacity-60">: float = 0.0</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">optim<span class="opacity-60">: typing.Union[transformers.training_args.OptimizerNames, str] = 'adamw_torch_fused'</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">optim_args<span class="opacity-60">: typing.Optional[str] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">adafactor<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">group_by_length<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">length_column_name<span class="opacity-60">: str = 'length'</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">report_to<span class="opacity-60">: typing.Union[NoneType, str, list[str]] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">project<span class="opacity-60">: str = 'huggingface'</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">trackio_space_id<span class="opacity-60">: typing.Optional[str] = 'trackio'</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">ddp_find_unused_parameters<span class="opacity-60">: typing.Optional[bool] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">ddp_bucket_cap_mb<span class="opacity-60">: typing.Optional[int] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">ddp_broadcast_buffers<span class="opacity-60">: typing.Optional[bool] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">dataloader_pin_memory<span class="opacity-60">: bool = True</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">dataloader_persistent_workers<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">skip_memory_metrics<span class="opacity-60">: bool = True</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">use_legacy_prediction_loop<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">push_to_hub<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">resume_from_checkpoint<span class="opacity-60">: typing.Optional[str] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">hub_model_id<span class="opacity-60">: typing.Optional[str] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">hub_strategy<span class="opacity-60">: typing.Union[transformers.trainer_utils.HubStrategy, str] = 'every_save'</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">hub_token<span class="opacity-60">: typing.Optional[str] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">hub_private_repo<span class="opacity-60">: typing.Optional[bool] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">hub_always_push<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">hub_revision<span class="opacity-60">: typing.Optional[str] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">gradient_checkpointing<span class="opacity-60">: bool = True</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">gradient_checkpointing_kwargs<span class="opacity-60">: typing.Union[dict[str, typing.Any], str, NoneType] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">include_inputs_for_metrics<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">include_for_metrics<span class="opacity-60">: list = <factory></span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">eval_do_concat_batches<span class="opacity-60">: bool = True</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">fp16_backend<span class="opacity-60">: str = 'auto'</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">push_to_hub_model_id<span class="opacity-60">: typing.Optional[str] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">push_to_hub_organization<span class="opacity-60">: typing.Optional[str] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">push_to_hub_token<span class="opacity-60">: typing.Optional[str] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">mp_parameters<span class="opacity-60">: str = ''</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">auto_find_batch_size<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">full_determinism<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">torchdynamo<span class="opacity-60">: typing.Optional[str] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">ray_scope<span class="opacity-60">: typing.Optional[str] = 'last'</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">ddp_timeout<span class="opacity-60">: int = 1800</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">torch_compile<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">torch_compile_backend<span class="opacity-60">: typing.Optional[str] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">torch_compile_mode<span class="opacity-60">: typing.Optional[str] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">include_tokens_per_second<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">include_num_input_tokens_seen<span class="opacity-60">: typing.Union[str, bool] = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">neftune_noise_alpha<span class="opacity-60">: typing.Optional[float] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">optim_target_modules<span class="opacity-60">: typing.Union[NoneType, str, list[str]] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">batch_eval_metrics<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">eval_on_start<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">use_liger_kernel<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">liger_kernel_config<span class="opacity-60">: typing.Optional[dict[str, bool]] = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">eval_use_gather_object<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">average_tokens_across_devices<span class="opacity-60">: bool = True</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">model_init_kwargs<span class="opacity-60">: dict | str | None = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">disable_dropout<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">cast_lm_head_to_fp32<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">max_prompt_length<span class="opacity-60">: int | None = 512</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">num_generations<span class="opacity-60">: int | None = 8</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">max_completion_length<span class="opacity-60">: int | None = 256</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">ds3_gather_for_generation<span class="opacity-60">: bool = True</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">shuffle_dataset<span class="opacity-60">: bool | None = True</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">generation_batch_size<span class="opacity-60">: int | None = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">steps_per_generation<span class="opacity-60">: int | None = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">temperature<span class="opacity-60">: float = 1.0</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">top_p<span class="opacity-60">: float = 1.0</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">top_k<span class="opacity-60">: int | None = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">min_p<span class="opacity-60">: float | None = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">generation_kwargs<span class="opacity-60">: dict | None = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">chat_template_kwargs<span class="opacity-60">: dict | None = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">repetition_penalty<span class="opacity-60">: float = 1.0</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">use_transformers_paged<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">cache_implementation<span class="opacity-60">: str | None = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">use_vllm<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">vllm_mode<span class="opacity-60">: str = 'server'</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">vllm_model_impl<span class="opacity-60">: str = 'vllm'</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">vllm_enable_sleep_mode<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">vllm_guided_decoding_regex<span class="opacity-60">: str | None = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">vllm_server_base_url<span class="opacity-60">: str | None = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">vllm_server_host<span class="opacity-60">: str = '0.0.0.0'</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">vllm_server_port<span class="opacity-60">: int = 8000</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">vllm_server_timeout<span class="opacity-60">: float = 240.0</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">vllm_gpu_memory_utilization<span class="opacity-60">: float = 0.3</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">vllm_tensor_parallel_size<span class="opacity-60">: int = 1</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">beta<span class="opacity-60">: float = 0.0</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">num_iterations<span class="opacity-60">: int = 1</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">epsilon<span class="opacity-60">: float = 0.2</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">delta<span class="opacity-60">: float | None = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">epsilon_high<span class="opacity-60">: float | None = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">importance_sampling_level<span class="opacity-60">: str = 'token'</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">reward_weights<span class="opacity-60">: list[float] | None = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">scale_rewards<span class="opacity-60">: str = 'group'</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">loss_type<span class="opacity-60">: str = 'dapo'</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">mask_truncated_completions<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">sync_ref_model<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">ref_model_mixup_alpha<span class="opacity-60">: float = 0.6</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">ref_model_sync_steps<span class="opacity-60">: int = 512</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">top_entropy_quantile<span class="opacity-60">: float = 1.0</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">use_liger_loss<span class="opacity-60">: bool = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">vllm_importance_sampling_correction<span class="opacity-60">: bool = True</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">vllm_importance_sampling_cap<span class="opacity-60">: float = 2.0</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">log_completions<span class="opacity-60">: bool = False</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">num_completions_to_print<span class="opacity-60">: int | None = None</span></span> </span><span class="comma cursor-default"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">wandb_log_unique_prompts<span class="opacity-60">: bool | None = False</span></span> </span> <span data-svelte-h="svelte-1jq0pl7">)</span> </p> <div class="!mb-10 relative docstring-details "> <p class="flex items-center font-semibold">Parameters that control the model and reference model <span class="flex-auto border-t-2 ml-3"></span></p> <ul class="px-2"><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.model_init_kwargs" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.model_init_kwargs"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>model_init_kwargs</strong> (<code>str</code>, <code>dict[str, Any]</code>, <em>optional</em>) — | |
| Keyword arguments for <a href="https://huggingface.co/docs/transformers/main/en/model_doc/auto#transformers.AutoModelForCausalLM.from_pretrained" rel="nofollow">from_pretrained</a>, used when the <code>model</code> | |
| argument of the <a href="/docs/trl/pr_4331/en/grpo_trainer#trl.GRPOTrainer">GRPOTrainer</a> is provided as a string.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.disable_dropout" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.disable_dropout"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>disable_dropout</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>False</code>) — | |
| Whether to disable dropout in the model. This is useful for training with a reference model, as it prevents | |
| the model from generating different logprobs for the same input.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.cast_lm_head_to_fp32" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.cast_lm_head_to_fp32"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>cast_lm_head_to_fp32</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>False</code>) — | |
| Whether to cast the language modeling head of the policy and reference models to float32. As recommended by | |
| the <a href="https://huggingface.co/papers/2510.13786" rel="nofollow">ScaleRL</a> recipe. This flag is only supported when the model | |
| has untied word embedding and language modeling head layers i.e. <code>tie_word_embeddings</code> in the model config | |
| is False.<!-- HTML_TAG_END --> </span></span> </li> </ul><p class="flex items-center font-semibold">Parameters that control the data preprocessing <span class="flex-auto border-t-2 ml-3"></span></p> <ul class="px-2"><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.remove_unused_columns" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.remove_unused_columns"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>remove_unused_columns</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>False</code>) — | |
| Whether to only keep the column <code>"prompt"</code> in the dataset. If you use a custom reward function that | |
| requires any column other than <code>"prompts"</code> and <code>"completions"</code>, you should keep this to <code>False</code>.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.max_prompt_length" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.max_prompt_length"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>max_prompt_length</strong> (<code>int</code> or <code>None</code>, <em>optional</em>, defaults to <code>512</code>) — | |
| Maximum length of the prompt. If the prompt is longer than this value, it will be truncated left.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.num_generations" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.num_generations"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>num_generations</strong> (<code>int</code> or <code>None</code>, <em>optional</em>, defaults to <code>8</code>) — | |
| Number of generations per prompt to sample. The effective batch size (num_processes * per_device_batch_size | |
| <ul> | |
| <li>gradient_accumulation_steps) must be evenly divisible by this value.</li> | |
| </ul><!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.max_completion_length" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.max_completion_length"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>max_completion_length</strong> (<code>int</code> or <code>None</code>, <em>optional</em>, defaults to <code>256</code>) — | |
| Maximum length of the generated completion.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.ds3_gather_for_generation" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.ds3_gather_for_generation"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>ds3_gather_for_generation</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>True</code>) — | |
| This setting applies to DeepSpeed ZeRO-3. If enabled, the policy model weights are gathered for generation, | |
| improving generation speed. However, disabling this option allows training models that exceed the VRAM | |
| capacity of a single GPU, albeit at the cost of slower generation. Disabling this option is not compatible | |
| with vLLM generation.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.shuffle_dataset" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.shuffle_dataset"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>shuffle_dataset</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>True</code>) — | |
| Whether to shuffle the training dataset.<!-- HTML_TAG_END --> </span></span> </li> </ul><p class="flex items-center font-semibold">Parameters that control generation <span class="flex-auto border-t-2 ml-3"></span></p> <ul class="px-2"><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.generation_batch_size" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.generation_batch_size"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>generation_batch_size</strong> — (<code>int</code>, <em>optional</em>): | |
| Batch size to use for generation. If <code>None</code>, it defaults to the effective training batch size: | |
| <code>per_device_train_batch_size * num_processes * steps_per_generation</code>. In other words, there is one | |
| generation batch processed per optimization step. Mutually exclusive with <code>steps_per_generation</code>.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.steps_per_generation" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.steps_per_generation"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>steps_per_generation</strong> — (<code>int</code>, <em>optional</em>): | |
| Number of steps per generation. If <code>None</code>, it defaults to <code>gradient_accumulation_steps</code>. Mutually exclusive | |
| with <code>generation_batch_size</code>.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.temperature" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.temperature"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>temperature</strong> (<code>float</code>, defaults to <code>1.0</code>) — | |
| Temperature for sampling. The higher the temperature, the more random the completions.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.top_p" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.top_p"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>top_p</strong> (<code>float</code>, <em>optional</em>, defaults to <code>1.0</code>) — | |
| Float that controls the cumulative probability of the top tokens to consider. Must be in (0, 1]. Set to | |
| <code>1.0</code> to consider all tokens.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.top_k" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.top_k"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>top_k</strong> (<code>int</code>, <em>optional</em>) — | |
| Number of highest probability vocabulary tokens to keep for top-k-filtering. If <code>None</code>, top-k-filtering is | |
| disabled and all tokens are considered.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.min_p" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.min_p"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>min_p</strong> (<code>float</code>, <em>optional</em>) — | |
| Minimum token probability, which will be scaled by the probability of the most likely token. It must be a | |
| value between <code>0.0</code> and <code>1.0</code>. Typical values are in the <code>0.01-0.2</code> range.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.generation_kwargs" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.generation_kwargs"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>generation_kwargs</strong> (<code>dict[str, Any]</code>, <em>optional</em>) — | |
| Additional keyword arguments to pass to <a href="https://huggingface.co/docs/transformers/main/en/main_classes/text_generation#transformers.GenerationConfig" rel="nofollow">GenerationConfig</a> (if using transformers) or | |
| <code>SamplingParams</code> (if using vLLM) when sampling completions. This can be used to further customize the | |
| generation behavior, such as setting <code>suppress_tokens</code>, <code>num_beams</code>, etc. If it contains keys that conflict | |
| with the other generation parameters (like <code>min_p</code>, <code>top_p</code>, etc.), they will override them.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.chat_template_kwargs" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.chat_template_kwargs"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>chat_template_kwargs</strong> (<code>dict[str, Any]</code>, <em>optional</em>) — | |
| Additional keyword arguments to pass to the <code>apply_chat_template</code> function when generating completions.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.repetition_penalty" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.repetition_penalty"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>repetition_penalty</strong> (<code>float</code>, <em>optional</em>, defaults to <code>1.0</code>) — | |
| Float that penalizes new tokens based on whether they appear in the prompt and the generated text so far. | |
| Values > <code>1.0</code> encourage the model to use new tokens, while values < <code>1.0</code> encourage the model to repeat | |
| tokens.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.use_transformers_paged" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.use_transformers_paged"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>use_transformers_paged</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>False</code>) — | |
| Whether to use the <code>transformers</code> paged implementation for generation. If set to <code>True</code>, the <code>transformers</code> | |
| paged implementation will be used for generation instead of the default padded implementation. This | |
| parameter is only effective when <code>use_vllm</code> is set to <code>False</code>.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.cache_implementation" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.cache_implementation"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>cache_implementation</strong> (<code>str</code>, <em>optional</em>) — | |
| Implementation of the cache method for faster generation when <code>use_vllm</code> is set to <code>False</code>.<!-- HTML_TAG_END --> </span></span> </li> </ul><p class="flex items-center font-semibold">Parameters that control generation acceleration powered by vLLM <span class="flex-auto border-t-2 ml-3"></span></p> <ul class="px-2"><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.use_vllm" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.use_vllm"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>use_vllm</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>False</code>) — | |
| Whether to use vLLM for generating completions. If set to <code>True</code>, the trainer will use vLLM for generation | |
| instead of the default model.generate(). Requires <code>vllm</code> to be installed.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.vllm_mode" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.vllm_mode"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>vllm_mode</strong> (<code>str</code>, <em>optional</em>, defaults to <code>"server"</code>) — | |
| Mode to use for vLLM integration when <code>use_vllm</code> is set to <code>True</code>. Must be one of <code>"server"</code> or | |
| <code>"colocate"</code>.</p> | |
| <ul> | |
| <li><code>"server"</code>: The trainer will send generation requests to a separate vLLM server. Make sure a TRL vLLM | |
| server is running (start with <code>trl vllm-serve</code>).</li> | |
| <li><code>"colocate"</code>: vLLM will run in the same process and share the training GPUs. This avoids the need for a | |
| separate server but may cause resource contention with training.</li> | |
| </ul><!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.vllm_model_impl" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.vllm_model_impl"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>vllm_model_impl</strong> (<code>str</code>, <em>optional</em>, defaults to <code>"vllm"</code>) — | |
| Model implementation to use for vLLM. Must be one of <code>"transformers"</code> or <code>"vllm"</code>. <code>"transformers"</code>: Use | |
| the <code>transformers</code> backend for model implementation. <code>"vllm"</code>: Use the <code>vllm</code> library for model | |
| implementation.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.vllm_guided_decoding_regex" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.vllm_guided_decoding_regex"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>vllm_guided_decoding_regex</strong> (<code>str</code>, <em>optional</em>) — | |
| Regex for vLLM guided decoding. If <code>None</code> (default), guided decoding is disabled.<!-- HTML_TAG_END --> </span></span> </li> </ul><p class="flex items-center font-semibold">Parameters that control the vLLM server (only used when `vllm_mode` is `"server"`) <span class="flex-auto border-t-2 ml-3"></span></p> <ul class="px-2"><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.vllm_server_base_url" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.vllm_server_base_url"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>vllm_server_base_url</strong> (<code>str</code>, <em>optional</em>) — | |
| Base URL for the vLLM server (e.g., <code>"http://localhost:8000"</code>). If provided, <code>vllm_server_host</code> and | |
| <code>vllm_server_port</code> are ignored.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.vllm_server_host" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.vllm_server_host"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>vllm_server_host</strong> (<code>str</code>, <em>optional</em>, defaults to <code>"0.0.0.0"</code>) — | |
| Host of the vLLM server to connect to. Ignored if <code>vllm_server_base_url</code> is provided.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.vllm_server_port" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.vllm_server_port"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>vllm_server_port</strong> (<code>int</code>, <em>optional</em>, defaults to <code>8000</code>) — | |
| Port of the vLLM server to connect to. Ignored if <code>vllm_server_base_url</code> is provided.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.vllm_server_timeout" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.vllm_server_timeout"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>vllm_server_timeout</strong> (<code>float</code>, <em>optional</em>, defaults to <code>240.0</code>) — | |
| Total timeout duration in seconds to wait for the vLLM server to be up. If the server is not up after the | |
| timeout, a <code>ConnectionError</code> is raised.<!-- HTML_TAG_END --> </span></span> </li> </ul><p class="flex items-center font-semibold">Parameters that control colocated vLLM execution (only used when `vllm_mode` is `"colocate"`) <span class="flex-auto border-t-2 ml-3"></span></p> <ul class="px-2"><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.vllm_gpu_memory_utilization" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.vllm_gpu_memory_utilization"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>vllm_gpu_memory_utilization</strong> (<code>float</code>, <em>optional</em>, defaults to <code>0.3</code>) — | |
| Control the GPU memory utilization for vLLM. This setting only applies when <code>vllm_mode</code> is set to | |
| <code>"colocate"</code>. If you are using <code>vllm_mode="server"</code>, this parameter must be passed separately when | |
| launching the vLLM server via the <code>--vllm_gpu_memory_utilization</code> flag.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.vllm_tensor_parallel_size" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.vllm_tensor_parallel_size"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>vllm_tensor_parallel_size</strong> (<code>int</code>, <em>optional</em>, defaults to <code>1</code>) — | |
| Control the tensor parallel size for vLLM. This setting only applies when <code>vllm_mode</code> is set to | |
| <code>"colocate"</code>. If you are using <code>vllm_mode="server"</code>, this parameter must be passed separately when | |
| launching the vLLM server via the <code>--vllm_tensor_parallel_size</code> flag.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.vllm_enable_sleep_mode" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.vllm_enable_sleep_mode"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>vllm_enable_sleep_mode</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>False</code>) — | |
| Whether to enable sleep mode for vLLM. If <code>True</code>, vLLM will sleep during the optimization step and woken | |
| for weight sync and generation.<!-- HTML_TAG_END --> </span></span> </li> </ul><p class="flex items-center font-semibold">Parameters that control the training <span class="flex-auto border-t-2 ml-3"></span></p> <ul class="px-2"><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.beta" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.beta"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>beta</strong> (<code>float</code>, <em>optional</em>, defaults to <code>0.0</code>) — | |
| KL coefficient. If <code>0.0</code> (default), the reference model is not loaded, reducing memory usage and improving | |
| training speed.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.num_iterations" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.num_iterations"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>num_iterations</strong> (<code>int</code>, <em>optional</em>, defaults to <code>1</code>) — | |
| Number of iterations per batch (denoted as μ in the algorithm).<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.epsilon" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.epsilon"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>epsilon</strong> (<code>float</code>, <em>optional</em>, defaults to <code>0.2</code>) — | |
| Epsilon value for clipping.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.delta" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.delta"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>delta</strong> (<code>float</code>, <em>optional</em>) — | |
| Enables the upper clipping bound in two-sided GRPO loss when set to a float. If <code>None</code> (default), standard | |
| GRPO clipping is used. Recommended to be greater than <code>1 + ε</code> when enabled. This method is introduced in | |
| the <a href="https://huggingface.co/papers/2505.07291" rel="nofollow">INTELLECT-2 tech report</a>.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.epsilon_high" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.epsilon_high"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>epsilon_high</strong> (<code>float</code>, <em>optional</em>) — | |
| Upper-bound epsilon value for clipping. If not specified, it defaults to the same value as the lower-bound | |
| specified in argument <code>epsilon</code>. Paper <a href="https://huggingface.co/papers/2503.14476" rel="nofollow">DAPO</a> recommends <code>0.28</code>.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.importance_sampling_level" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.importance_sampling_level"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>importance_sampling_level</strong> (<code>str</code>, <em>optional</em>, defaults to <code>"token"</code>) — | |
| Controls whether importance sampling ratios are computed at the <code>"token"</code> or <code>"sequence"</code> level. <code>"token"</code> | |
| keeps the raw per-token log-probability ratios (one weight per token). <code>"sequence"</code> averages the | |
| log-probability ratios across valid tokens to produce a single ratio per sequence. The <a href="https://huggingface.co/papers/2507.18071" rel="nofollow">GSPO | |
| paper</a> shows that sequence-level sampling often yields more | |
| stable training and better alignment with sequence-level rewards.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.reward_weights" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.reward_weights"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>reward_weights</strong> (<code>list[float]</code>, <em>optional</em>) — | |
| Weights for each reward function. Must match the number of reward functions. If <code>None</code>, all rewards are | |
| weighted equally with weight <code>1.0</code>.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.scale_rewards" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.scale_rewards"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>scale_rewards</strong> (<code>str</code> or <code>bool</code>, <em>optional</em>, defaults to <code>"group"</code>) — | |
| Specifies the scaling strategy for rewards. Supported values are:</p> | |
| <ul> | |
| <li><code>True</code> or <code>"group"</code> (default): rewards are scaled by the standard deviation within each group, ensuring | |
| unit variance within a group.</li> | |
| <li><code>"batch"</code>: rewards are scaled by the standard deviation across the entire batch, as recommended in the | |
| <a href="https://huggingface.co/papers/2508.08221" rel="nofollow">PPO Lite paper</a>.</li> | |
| <li><code>False</code> or <code>"none"</code>: no scaling is applied. The <a href="https://huggingface.co/papers/2503.20783" rel="nofollow">Dr. GRPO | |
| paper</a> recommends not scaling rewards, as scaling by the | |
| standard deviation introduces a question-level difficulty bias.</li> | |
| </ul><!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.loss_type" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.loss_type"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>loss_type</strong> (<code>str</code>, <em>optional</em>, defaults to <code>"dapo"</code>) — | |
| Specifies the loss formulation to use. Supported values are:</p> | |
| <ul> | |
| <li><code>"grpo"</code>: Aggregates token-level losses by normalizing over sequence length. Not recommended due to | |
| length bias—this approach tends to prefer shorter completions with positive advantages and longer ones | |
| with negative advantages.</li> | |
| <li><code>"dr_grpo"</code>: Aggregates token-level losses by normalizing with a global constant. This method was | |
| introduced in the <a href="https://huggingface.co/papers/2503.20783" rel="nofollow">Dr. GRPO paper</a> to eliminate length bias. | |
| The value of the constant corresponds to <code>max_completion_length</code>.</li> | |
| <li><code>"dapo"</code> (default): Aggregates token-level losses by normalizing with the number of active token in the | |
| global accumulated batch. This method was introduced in the <a href="https://huggingface.co/papers/2503.14476" rel="nofollow">DAPO | |
| paper</a> to eliminate length bias.</li> | |
| <li><code>"bnpo"</code>: Aggregates token-level losses by normalizing with the number of active token in the local | |
| batch. Note that normalization is performed over the local batch only, so results may slightly vary | |
| depending on the local batch size, despite a constant effective batch size. When using | |
| <code>per_device_train_batch_size==1</code>, the loss is equivalent to the GRPO loss.</li> | |
| </ul><!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.mask_truncated_completions" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.mask_truncated_completions"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>mask_truncated_completions</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>False</code>) — | |
| When enabled, truncated completions are excluded from the loss calculation, preventing them from being | |
| incorrectly penalized and introducing noise during training. According to the | |
| <a href="https://huggingface.co/papers/2503.14476" rel="nofollow">DAPO</a> paper, this is a good practice for training stability.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.sync_ref_model" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.sync_ref_model"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>sync_ref_model</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>False</code>) — | |
| Whether to synchronize the reference model with the active model every <code>ref_model_sync_steps</code> steps, using | |
| the <code>ref_model_mixup_alpha</code> parameter. This synchronization originates from the | |
| <a href="https://huggingface.co/papers/2404.09656" rel="nofollow">TR-DPO</a> paper.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.ref_model_mixup_alpha" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.ref_model_mixup_alpha"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>ref_model_mixup_alpha</strong> (<code>float</code>, <em>optional</em>, defaults to <code>0.6</code>) — | |
| α parameter from the <a href="https://huggingface.co/papers/2404.09656" rel="nofollow">TR-DPO</a> paper, which controls the mix | |
| between the current policy and the previous reference policy during updates. The reference policy is | |
| updated according to the equation: <code>π_ref = α * π_θ + (1 - α) * π_ref_prev</code>. To use this parameter, you | |
| must set <code>sync_ref_model=True</code>.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.ref_model_sync_steps" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.ref_model_sync_steps"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>ref_model_sync_steps</strong> (<code>int</code>, <em>optional</em>, defaults to <code>512</code>) — | |
| τ parameter from the <a href="https://huggingface.co/papers/2404.09656" rel="nofollow">TR-DPO</a> paper, which determines how | |
| frequently the current policy is synchronized with the reference policy. To use this parameter, you must | |
| set <code>sync_ref_model=True</code>.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.top_entropy_quantile" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.top_entropy_quantile"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>top_entropy_quantile</strong> (<code>float</code>, <em>optional</em>, defaults to <code>1.0</code>) — | |
| ρ parameter from <a href="https://huggingface.co/papers/2506.01939" rel="nofollow">Beyond the 80/20 Rule</a>. Keeps in the policy | |
| loss term only the top-ρ quantile of tokens by entropy of the probability distribution at each sequence | |
| position, improving results. Range: <code>[0.0-1.0]</code>. A value of <code>0.0</code> masks all but the highest entropy token; | |
| <code>1.0</code> keeps all tokens. The paper recommends a value of <code>0.2</code>. If used with | |
| <code>mask_truncated_completions=True</code>, only tokens from non-truncated completions are considered.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.use_liger_loss" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.use_liger_loss"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>use_liger_loss</strong> (<code>bool</code>, <em>optional</em>) — | |
| Whether to use Liger loss.</p> | |
| <deprecated version="0.25.0"> | |
| <p>Parameter <code>use_liger_loss</code> is deprecated and will be removed in version 0.28.0. Use <code>use_liger_kernel</code> | |
| instead.</p> | |
| </deprecated><!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.vllm_importance_sampling_correction" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.vllm_importance_sampling_correction"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>vllm_importance_sampling_correction</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>True</code>) — | |
| Whether to apply Truncated Importance Sampling (TIS) between vLLM completion logprobs and recomputed | |
| logprobs. <a href="https://fengyao.notion.site/off-policy-rl" rel="nofollow">Your Efficient RL Framework Secretly Brings You Off-Policy RL | |
| Training</a> highlights that using a separate generation framework | |
| (such as vLLM) can introduce off-policy effects due to subtle implementation differences between generation | |
| and training backends. TIS is proposed as a remedy for this issue.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.vllm_importance_sampling_cap" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.vllm_importance_sampling_cap"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>vllm_importance_sampling_cap</strong> (<code>float</code>, <em>optional</em>, defaults to <code>2.0</code>) — | |
| Truncation parameter C for Truncated Importance Sampling (TIS). This sets an upper bound on the importance | |
| sampling ratio, improving training stability.<!-- HTML_TAG_END --> </span></span> </li> </ul><p class="flex items-center font-semibold">Parameters that control the logging <span class="flex-auto border-t-2 ml-3"></span></p> <ul class="px-2"><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.log_completions" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.log_completions"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>log_completions</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>False</code>) — | |
| Whether to log a sample of (prompt, completion) pairs every <code>logging_steps</code> steps. If <code>rich</code> is installed, | |
| it prints the sample. If <code>wandb</code> and/or <code>trackio</code> logging is enabled, it logs it to <code>wandb</code> and/or | |
| <code>trackio</code>.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.num_completions_to_print" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.num_completions_to_print"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>num_completions_to_print</strong> (<code>int</code>, <em>optional</em>) — | |
| Number of completions to print with <code>rich</code>. If <code>None</code>, all completions are logged.<!-- HTML_TAG_END --> </span></span> </li><li class="text-base !pl-4 my-3 rounded "><span class="group flex space-x-1.5 items-start"><a id="trl.GRPOConfig.wandb_log_unique_prompts" class="header-link block pr-0.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#trl.GRPOConfig.wandb_log_unique_prompts"><span><svg class="text-smd" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span><!-- HTML_TAG_START --><strong>wandb_log_unique_prompts</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>False</code>) — | |
| Whether to log unique prompts in wandb. If <code>True</code>, only unique prompts are logged. If <code>False</code>, all prompts | |
| are logged.<!-- HTML_TAG_END --> </span></span> </li> </ul> </div></div> <p data-svelte-h="svelte-1xkpnbh">Configuration class for the <a href="/docs/trl/pr_4331/en/grpo_trainer#trl.GRPOTrainer">GRPOTrainer</a>.</p> <p data-svelte-h="svelte-13s1zc4">This class includes only the parameters that are specific to GRPO training. For a full list of training arguments, | |
| please refer to the <a href="https://huggingface.co/docs/transformers/main/en/main_classes/trainer#transformers.TrainingArguments" rel="nofollow">TrainingArguments</a> documentation. Note that default values in this class may | |
| differ from those in <a href="https://huggingface.co/docs/transformers/main/en/main_classes/trainer#transformers.TrainingArguments" rel="nofollow">TrainingArguments</a>.</p> <p data-svelte-h="svelte-ekuf1t">Using <a href="https://huggingface.co/docs/transformers/main/en/internal/trainer_utils#transformers.HfArgumentParser" rel="nofollow">HfArgumentParser</a> we can turn this class into | |
| <a href="https://docs.python.org/3/library/argparse#module-argparse" rel="nofollow">argparse</a> arguments that can be specified on the | |
| command line.</p></div> <a class="!text-gray-400 !no-underline text-sm flex items-center not-prose mt-4" href="https://github.com/huggingface/trl/blob/main/docs/source/grpo_trainer.md" target="_blank"><svg class="mr-1" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M31,16l-7,7l-1.41-1.41L28.17,16l-5.58-5.59L24,9l7,7z"></path><path d="M1,16l7-7l1.41,1.41L3.83,16l5.58,5.59L8,23l-7-7z"></path><path d="M12.419,25.484L17.639,6.552l1.932,0.518L14.351,26.002z"></path></svg> <span data-svelte-h="svelte-zjs2n5"><span class="underline">Update</span> on GitHub</span></a> <p></p> | |
| <script> | |
| { | |
| __sveltekit_l5mord = { | |
| assets: "/docs/trl/pr_4331/en", | |
| base: "/docs/trl/pr_4331/en", | |
| env: {} | |
| }; | |
| const element = document.currentScript.parentElement; | |
| const data = [null,null]; | |
| Promise.all([ | |
| import("/docs/trl/pr_4331/en/_app/immutable/entry/start.6bbbc54b.js"), | |
| import("/docs/trl/pr_4331/en/_app/immutable/entry/app.b003256e.js") | |
| ]).then(([kit, app]) => { | |
| kit.start(app, element, { | |
| node_ids: [0, 19], | |
| data, | |
| form: null, | |
| error: null | |
| }); | |
| }); | |
| } | |
| </script> | |
Xet Storage Details
- Size:
- 427 kB
- Xet hash:
- 8dbd4a8574799f2d8b0259c10fb8598ba9f0ecce9489721357fee4228f3ebcd7
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.