Buckets:

hf-doc-build
/

doc-dev

Files

xet

hf-doc-build/doc-dev / course /pr_1069 /en /chapter12 /4.html

rtrm

28 days ago

download

raw

38.5 kB

	<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{"title":"Implementing GRPO in TRL","local":"implementing-grpo-in-trl","sections":[{"title":"Key Components","local":"key-components","sections":[{"title":"1. Dataset Format","local":"1-dataset-format","sections":[],"depth":3},{"title":"2. Reward Function","local":"2-reward-function","sections":[],"depth":3},{"title":"3. Training Configuration","local":"3-training-configuration","sections":[],"depth":3}],"depth":2},{"title":"Tips for Success","local":"tips-for-success","sections":[],"depth":2},{"title":"Reward Function Design","local":"reward-function-design","sections":[{"title":"1. Length-Based Rewards","local":"1-length-based-rewards","sections":[],"depth":3}],"depth":2},{"title":"2. Rule-Based Rewards for Verifiable Tasks","local":"2-rule-based-rewards-for-verifiable-tasks","sections":[],"depth":2},{"title":"3. Format-Based Rewards","local":"3-format-based-rewards","sections":[],"depth":2},{"title":"That’s it!","local":"thats-it","sections":[],"depth":2}],"depth":1}">
	<link href="/docs/course/pr_1069/en/_app/immutable/assets/0.e3b0c442.css" rel="modulepreload">
	<link rel="modulepreload" href="/docs/course/pr_1069/en/_app/immutable/entry/start.c5306bb2.js">
	<link rel="modulepreload" href="/docs/course/pr_1069/en/_app/immutable/chunks/scheduler.37c15a92.js">
	<link rel="modulepreload" href="/docs/course/pr_1069/en/_app/immutable/chunks/singletons.bc78d867.js">
	<link rel="modulepreload" href="/docs/course/pr_1069/en/_app/immutable/chunks/index.18351ede.js">
	<link rel="modulepreload" href="/docs/course/pr_1069/en/_app/immutable/chunks/paths.76894643.js">
	<link rel="modulepreload" href="/docs/course/pr_1069/en/_app/immutable/entry/app.4264f5f8.js">
	<link rel="modulepreload" href="/docs/course/pr_1069/en/_app/immutable/chunks/index.7cb9c9b8.js">
	<link rel="modulepreload" href="/docs/course/pr_1069/en/_app/immutable/nodes/0.f5347c47.js">
	<link rel="modulepreload" href="/docs/course/pr_1069/en/_app/immutable/chunks/each.e59479a4.js">
	<link rel="modulepreload" href="/docs/course/pr_1069/en/_app/immutable/nodes/32.8f441080.js">
	<link rel="modulepreload" href="/docs/course/pr_1069/en/_app/immutable/chunks/Tip.d10b3fc9.js">
	<link rel="modulepreload" href="/docs/course/pr_1069/en/_app/immutable/chunks/CodeBlock.abae2786.js">
	<link rel="modulepreload" href="/docs/course/pr_1069/en/_app/immutable/chunks/getInferenceSnippets.f9350a3f.js"><!-- HEAD_svelte-u9bgzb_START --><meta name="hf:doc:metadata" content="{"title":"Implementing GRPO in TRL","local":"implementing-grpo-in-trl","sections":[{"title":"Key Components","local":"key-components","sections":[{"title":"1. Dataset Format","local":"1-dataset-format","sections":[],"depth":3},{"title":"2. Reward Function","local":"2-reward-function","sections":[],"depth":3},{"title":"3. Training Configuration","local":"3-training-configuration","sections":[],"depth":3}],"depth":2},{"title":"Tips for Success","local":"tips-for-success","sections":[],"depth":2},{"title":"Reward Function Design","local":"reward-function-design","sections":[{"title":"1. Length-Based Rewards","local":"1-length-based-rewards","sections":[],"depth":3}],"depth":2},{"title":"2. Rule-Based Rewards for Verifiable Tasks","local":"2-rule-based-rewards-for-verifiable-tasks","sections":[],"depth":2},{"title":"3. Format-Based Rewards","local":"3-format-based-rewards","sections":[],"depth":2},{"title":"That’s it!","local":"thats-it","sections":[],"depth":2}],"depth":1}"><!-- HEAD_svelte-u9bgzb_END --> <p></p> <h1 class="relative group"><a id="implementing-grpo-in-trl" 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="#implementing-grpo-in-trl"><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>Implementing GRPO in TRL</span></h1> <p data-svelte-h="svelte-v9iq5e">In this page, we’ll learn how to implement Group Relative Policy Optimization (GRPO) using the Transformer Reinforcement Learning (TRL) library. We’ll focus on practical implementation with minimal code.</p> <p data-svelte-h="svelte-iqdzey">We’ll explore the core concepts of GRPO as they are embodied in TRL’s GRPOTrainer, using snippets from the official TRL documentation to guide us.</p> <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><p data-svelte-h="svelte-14oapf3">This chapter is aimed at TRL beginners. If you are already familiar with TRL, you might want to also check out the <a href="https://github.com/huggingface/open-r1/blob/main/src/open_r1/grpo.py" rel="nofollow">Open R1 implementation</a> of GRPO.</p></div> <p data-svelte-h="svelte-14hk95q">First, let’s remind ourselves of some of the important concepts of GRPO algorithm:</p> <ul data-svelte-h="svelte-1vd6wz6"><li>Group Formation: The model generates multiple completions for each prompt.</li> <li>Preference Learning: The model learns from a reward function that compares groups of completions.</li> <li>Training Configuration: The model uses a configuration to control the training process.</li></ul> <p data-svelte-h="svelte-pzoe2y">What do we need to do to implement GRPO?</p> <ul data-svelte-h="svelte-1hif2as"><li>Define a dataset of prompts.</li> <li>Define a reward function that takes a list of completions and returns a list of rewards.</li> <li>Configure the training process with a GRPOConfig.</li> <li>Train the model using the GRPOTrainer.</li></ul> <p data-svelte-h="svelte-ac9u4w">Here’s a minimal example to get started with GRPO training:</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, GRPOConfig
	<span class="hljs-keyword">from</span> datasets <span class="hljs-keyword">import</span> load_dataset

	<span class="hljs-comment"># 1. Load your dataset</span>
	dataset = load_dataset(<span class="hljs-string">"your_dataset"</span>, split=<span class="hljs-string">"train"</span>)


	<span class="hljs-comment"># 2. Define a simple reward function</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-string">"""Example: Reward longer completions"""</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]


	<span class="hljs-comment"># 3. Configure training</span>
	training_args = GRPOConfig(
	output_dir=<span class="hljs-string">"output"</span>,
	num_train_epochs=<span class="hljs-number">3</span>,
	per_device_train_batch_size=<span class="hljs-number">4</span>,
	gradient_accumulation_steps=<span class="hljs-number">2</span>,
	logging_steps=<span class="hljs-number">10</span>,
	)

	<span class="hljs-comment"># 4. Initialize and train</span>
	trainer = GRPOTrainer(
	model=<span class="hljs-string">"your_model"</span>, <span class="hljs-comment"># e.g. "Qwen/Qwen2-0.5B-Instruct"</span>
	args=training_args,
	train_dataset=dataset,
	reward_funcs=reward_func,
	)
	trainer.train()<!-- HTML_TAG_END --></pre></div> <h2 class="relative group"><a id="key-components" 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="#key-components"><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>Key Components</span></h2> <h3 class="relative group"><a id="1-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="#1-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>1. Dataset Format</span></h3> <p data-svelte-h="svelte-1pnmwnq">Your dataset should contain prompts that the model will respond to. The GRPO trainer will generate multiple completions for each prompt and use the reward function to compare them.</p> <h3 class="relative group"><a id="2-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="#2-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>2. Reward Function</span></h3> <p data-svelte-h="svelte-ths948">The reward function is crucial - it determines how the model learns. Here are two practical examples:</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"># Example 1: Reward based on completion length</span>
	<span class="hljs-keyword">def</span> <span class="hljs-title function_">reward_length</span>(<span class="hljs-params">completions, **kwargs</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]


	<span class="hljs-comment"># Example 2: Reward based on matching a pattern</span>
	<span class="hljs-keyword">import</span> re


	<span class="hljs-keyword">def</span> <span class="hljs-title function_">reward_format</span>(<span class="hljs-params">completions, **kwargs</span>):
	pattern = <span class="hljs-string">r"^<think>.?</think><answer>.?</answer>$"</span>
	<span class="hljs-keyword">return</span> [<span class="hljs-number">1.0</span> <span class="hljs-keyword">if</span> re.<span class="hljs-keyword">match</span>(pattern, c) <span class="hljs-keyword">else</span> <span class="hljs-number">0.0</span> <span class="hljs-keyword">for</span> c <span class="hljs-keyword">in</span> completions]<!-- HTML_TAG_END --></pre></div> <h3 class="relative group"><a id="3-training-configuration" 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="#3-training-configuration"><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>3. Training Configuration</span></h3> <p data-svelte-h="svelte-1uukcnx">Key parameters to consider in <code>GRPOConfig</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 -->training_args = GRPOConfig(
	<span class="hljs-comment"># Essential parameters</span>
	output_dir=<span class="hljs-string">"output"</span>,
	num_train_epochs=<span class="hljs-number">3</span>,
	num_generation=<span class="hljs-number">4</span>, <span class="hljs-comment"># Number of completions to generate for each prompt</span>
	per_device_train_batch_size=<span class="hljs-number">4</span>, <span class="hljs-comment"># We want to get all generations in one device batch</span>
	<span class="hljs-comment"># Optional but useful</span>
	gradient_accumulation_steps=<span class="hljs-number">2</span>,
	learning_rate=<span class="hljs-number">1e-5</span>,
	logging_steps=<span class="hljs-number">10</span>,
	<span class="hljs-comment"># GRPO specific (optional)</span>
	use_vllm=<span class="hljs-literal">True</span>, <span class="hljs-comment"># Speed up generation</span>
	)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1tdjaxe">The <code>num_generation</code> parameter is particularly important for GRPO as it defines the group size - how many different completions the model will generate for each prompt. This is a key differentiator from other RL methods:</p> <ul data-svelte-h="svelte-1b1t69f"><li>Too small (e.g., 2-3): May not provide enough diversity for meaningful comparisons</li> <li>Recommended (4-16): Provides good balance between diversity and computational efficiency</li> <li>Larger values: May improve learning but significantly increases computational cost</li></ul> <p data-svelte-h="svelte-7howml">The group size should be chosen based on your computational resources and the complexity of your task. For simple tasks, smaller groups (4-8) may be sufficient, while more complex reasoning tasks might benefit from larger groups (8-16).</p> <h2 class="relative group"><a id="tips-for-success" 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="#tips-for-success"><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>Tips for Success</span></h2> <ol data-svelte-h="svelte-4ukx5l"><li><strong>Memory Management</strong>: Adjust <code>per_device_train_batch_size</code> and <code>gradient_accumulation_steps</code> based on your GPU memory.</li> <li><strong>Speed</strong>: Enable <code>use_vllm=True</code> for faster generation if your model is supported.</li> <li><strong>Monitoring</strong>: Watch the logged metrics during training:<ul><li><code>reward</code>: Average reward across completions</li> <li><code>reward_std</code>: Standard deviation within reward groups</li> <li><code>kl</code>: KL divergence from reference model</li></ul></li></ol> <h2 class="relative group"><a id="reward-function-design" 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="#reward-function-design"><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>Reward Function Design</span></h2> <p data-svelte-h="svelte-5lvuy8">The DeepSeek R1 paper demonstrates several effective approaches to reward function design that you can adapt for your own GRPO implementation:</p> <h3 class="relative group"><a id="1-length-based-rewards" 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="#1-length-based-rewards"><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>1. Length-Based Rewards</span></h3> <p data-svelte-h="svelte-jhf03w">One of the easiest rewards to implement is a length-based reward. You can reward 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_len</span>(<span class="hljs-params">completions, **kwargs</span>):
	ideal_length = <span class="hljs-number">20</span>
	<span class="hljs-keyword">return</span> [-<span class="hljs-built_in">abs</span>(ideal_length - <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-jlmna3">This reward function penalizes completions that are too short or too long, encouraging the model to generate completions that are close to the ideal length of 20 tokens.</p> <iframe src="https://marimo.app/gh/huggingface/notebooks/main/e?entrypoint=course%2Fen%2Fchapter13%2Fgrpo_length.py&embed=true&show-chrome=false" title="Marimo Notebook" width="100%" height="800px" frameborder="0" allow="clipboard-write"></iframe> <h2 class="relative group"><a id="2-rule-based-rewards-for-verifiable-tasks" 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="#2-rule-based-rewards-for-verifiable-tasks"><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>2. Rule-Based Rewards for Verifiable Tasks</span></h2> <p data-svelte-h="svelte-mrpax">For tasks with objectively correct answers (like mathematics or coding), you can implement rule-based reward functions:</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_">problem_reward</span>(<span class="hljs-params">completions, answers, **kwargs</span>):
	<span class="hljs-string">"""Reward function for math problems with verifiable answers
	completions: list of completions to evaluate
	answers: list of answers to the problems from the dataset
	"""</span>

	rewards = []
	<span class="hljs-keyword">for</span> completion, correct_answer <span class="hljs-keyword">in</span> <span class="hljs-built_in">zip</span>(completions, answers):
	<span class="hljs-comment"># Extract the answer from the completion</span>
	<span class="hljs-keyword">try</span>:
	<span class="hljs-comment"># This is a simplified example - you'd need proper parsing</span>
	answer = extract_final_answer(completion)
	<span class="hljs-comment"># Binary reward: 1 for correct, 0 for incorrect</span>
	reward = <span class="hljs-number">1.0</span> <span class="hljs-keyword">if</span> answer == correct_answer <span class="hljs-keyword">else</span> <span class="hljs-number">0.0</span>
	rewards.append(reward)
	<span class="hljs-keyword">except</span>:
	<span class="hljs-comment"># If we can't parse an answer, give a low reward</span>
	rewards.append(<span class="hljs-number">0.0</span>)

	<span class="hljs-keyword">return</span> rewards<!-- HTML_TAG_END --></pre></div> <iframe src="https://marimo.app/gh/huggingface/notebooks/main/e?entrypoint=course%2Fen%2Fchapter13%2Fgrpo_math.py&embed=true&show-chrome=false" title="Marimo Notebook" width="100%" height="800px" frameborder="0" allow="clipboard-write"></iframe> <h2 class="relative group"><a id="3-format-based-rewards" 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="#3-format-based-rewards"><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>3. Format-Based Rewards</span></h2> <p data-svelte-h="svelte-18k4wdv">You can also reward proper formatting, which was important in the DeepSeek R1 training:</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_">format_reward</span>(<span class="hljs-params">completions, **kwargs</span>):
	<span class="hljs-string">"""Reward completions that follow the desired format"""</span>
	<span class="hljs-comment"># Example: Check if the completion follows a think-then-answer format</span>
	pattern = <span class="hljs-string">r"<think>(.?)</think>\s<answer>(.*?)</answer>"</span>

	rewards = []
	<span class="hljs-keyword">for</span> completion <span class="hljs-keyword">in</span> completions:
	<span class="hljs-keyword">match</span> = re.search(pattern, completion, re.DOTALL)
	<span class="hljs-keyword">if</span> <span class="hljs-keyword">match</span>:
	<span class="hljs-comment"># Check if there's substantial content in both sections</span>
	think_content = <span class="hljs-keyword">match</span>.group(<span class="hljs-number">1</span>).strip()
	answer_content = <span class="hljs-keyword">match</span>.group(<span class="hljs-number">2</span>).strip()

	<span class="hljs-keyword">if</span> <span class="hljs-built_in">len</span>(think_content) > <span class="hljs-number">20</span> <span class="hljs-keyword">and</span> <span class="hljs-built_in">len</span>(answer_content) > <span class="hljs-number">0</span>:
	rewards.append(<span class="hljs-number">1.0</span>)
	<span class="hljs-keyword">else</span>:
	rewards.append(
	<span class="hljs-number">0.5</span>
	) <span class="hljs-comment"># Partial reward for correct format but limited content</span>
	<span class="hljs-keyword">else</span>:
	rewards.append(<span class="hljs-number">0.0</span>) <span class="hljs-comment"># No reward for incorrect format</span>

	<span class="hljs-keyword">return</span> rewards<!-- HTML_TAG_END --></pre></div> <iframe src="https://marimo.app/gh/huggingface/notebooks/main/e?entrypoint=course%2Fen%2Fchapter13%2Fgrpo_format.py&embed=true&show-chrome=false" title="Marimo Notebook" width="100%" height="800px" frameborder="0" allow="clipboard-write"></iframe> <p data-svelte-h="svelte-fzl1fe">These examples demonstrate how you can implement reward functions inspired by the DeepSeek R1 training process, focusing on correctness, formatting, and combined signals.</p> <h2 class="relative group"><a id="thats-it" 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="#thats-it"><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>That’s it!</span></h2> <p data-svelte-h="svelte-chd5s6">In the next section, you will follow an exercise to implement GRPO in TRL.</p> <a class="!text-gray-400 !no-underline text-sm flex items-center not-prose mt-4" href="https://github.com/huggingface/course/blob/main/chapters/en/chapter12/4.mdx" target="_blank"><span data-svelte-h="svelte-1kd6by1"><</span> <span data-svelte-h="svelte-x0xyl0">></span> <span data-svelte-h="svelte-1dajgef"><span class="underline ml-1.5">Update</span> on GitHub</span></a> <p></p>

	<script>
	{
	__sveltekit_1y0degu = {
	assets: "/docs/course/pr_1069/en",
	base: "/docs/course/pr_1069/en",
	env: {}
	};

	const element = document.currentScript.parentElement;

	const data = [null,null];

	Promise.all([
	import("/docs/course/pr_1069/en/_app/immutable/entry/start.c5306bb2.js"),
	import("/docs/course/pr_1069/en/_app/immutable/entry/app.4264f5f8.js")
	]).then(([kit, app]) => {
	kit.start(app, element, {
	node_ids: [0, 32],
	data,
	form: null,
	error: null
	});
	});
	}
	</script>

Xet Storage Details

Size:: 38.5 kB
Xet hash:: 6137e9f148d457ccdac439a1fe9b5498daf6e625cda91c5e0df20d3d32db1ada

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.