Buckets:

hf-doc-build
/

doc-dev

Files

xet

hf-doc-build/doc-dev / course /pr_1213 /es /chapter12 /4.html

rtrm

3 months ago

download

raw

39 kB

	<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{"title":"Implementar GRPO en TRL","local":"implementar-grpo-en-trl","sections":[{"title":"Componentes clave","local":"componentes-clave","sections":[{"title":"1. Formato del dataset","local":"1-formato-del-dataset","sections":[],"depth":3},{"title":"2. Función de recompensa","local":"2-función-de-recompensa","sections":[],"depth":3},{"title":"3. Configuración del entrenamiento","local":"3-configuración-del-entrenamiento","sections":[],"depth":3}],"depth":2},{"title":"Consejos para tener éxito","local":"consejos-para-tener-éxito","sections":[],"depth":2},{"title":"Diseño de funciones de recompensa","local":"diseño-de-funciones-de-recompensa","sections":[{"title":"1. Recompensas basadas en longitud","local":"1-recompensas-basadas-en-longitud","sections":[],"depth":3}],"depth":2},{"title":"2. Recompensas basadas en reglas para tareas verificables","local":"2-recompensas-basadas-en-reglas-para-tareas-verificables","sections":[],"depth":2},{"title":"3. Recompensas basadas en formato","local":"3-recompensas-basadas-en-formato","sections":[],"depth":2},{"title":"¡Eso es todo!","local":"eso-es-todo","sections":[],"depth":2}],"depth":1}">
	<link href="/docs/course/pr_1213/es/_app/immutable/assets/0.e3b0c442.css" rel="modulepreload">
	<link rel="modulepreload" href="/docs/course/pr_1213/es/_app/immutable/entry/start.36d27295.js">
	<link rel="modulepreload" href="/docs/course/pr_1213/es/_app/immutable/chunks/scheduler.505acc25.js">
	<link rel="modulepreload" href="/docs/course/pr_1213/es/_app/immutable/chunks/singletons.6865fa96.js">
	<link rel="modulepreload" href="/docs/course/pr_1213/es/_app/immutable/chunks/index.001f95d5.js">
	<link rel="modulepreload" href="/docs/course/pr_1213/es/_app/immutable/chunks/paths.ec28c642.js">
	<link rel="modulepreload" href="/docs/course/pr_1213/es/_app/immutable/entry/app.3b43d7f3.js">
	<link rel="modulepreload" href="/docs/course/pr_1213/es/_app/immutable/chunks/preload-helper.8c2bab6b.js">
	<link rel="modulepreload" href="/docs/course/pr_1213/es/_app/immutable/chunks/index.e22abd30.js">
	<link rel="modulepreload" href="/docs/course/pr_1213/es/_app/immutable/nodes/0.e2c0ea78.js">
	<link rel="modulepreload" href="/docs/course/pr_1213/es/_app/immutable/chunks/each.e59479a4.js">
	<link rel="modulepreload" href="/docs/course/pr_1213/es/_app/immutable/nodes/32.cc31d760.js">
	<link rel="modulepreload" href="/docs/course/pr_1213/es/_app/immutable/chunks/MermaidChart.svelte_svelte_type_style_lang.a144e953.js">
	<link rel="modulepreload" href="/docs/course/pr_1213/es/_app/immutable/chunks/CodeBlock.f6688f67.js"><!-- HEAD_svelte-u9bgzb_START --><meta name="hf:doc:metadata" content="{"title":"Implementar GRPO en TRL","local":"implementar-grpo-en-trl","sections":[{"title":"Componentes clave","local":"componentes-clave","sections":[{"title":"1. Formato del dataset","local":"1-formato-del-dataset","sections":[],"depth":3},{"title":"2. Función de recompensa","local":"2-función-de-recompensa","sections":[],"depth":3},{"title":"3. Configuración del entrenamiento","local":"3-configuración-del-entrenamiento","sections":[],"depth":3}],"depth":2},{"title":"Consejos para tener éxito","local":"consejos-para-tener-éxito","sections":[],"depth":2},{"title":"Diseño de funciones de recompensa","local":"diseño-de-funciones-de-recompensa","sections":[{"title":"1. Recompensas basadas en longitud","local":"1-recompensas-basadas-en-longitud","sections":[],"depth":3}],"depth":2},{"title":"2. Recompensas basadas en reglas para tareas verificables","local":"2-recompensas-basadas-en-reglas-para-tareas-verificables","sections":[],"depth":2},{"title":"3. Recompensas basadas en formato","local":"3-recompensas-basadas-en-formato","sections":[],"depth":2},{"title":"¡Eso es todo!","local":"eso-es-todo","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 h-7 max-sm:h-7 px-2 max-sm:px-1.5 text-sm 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 hover:text-gray-800 dark:hover:text-gray-200"><svg class="sm:size-3.5 size-3" 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-7 max-sm:h-7 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 sm:size-3.5 size-3 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="implementar-grpo-en-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="#implementar-grpo-en-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>Implementar GRPO en TRL</span></h1> <p data-svelte-h="svelte-927fg4">En esta página aprenderemos a implementar Group Relative Policy Optimization (GRPO) usando la librería Transformer Reinforcement Learning (TRL). Nos centraremos en una implementación práctica con el mínimo código.</p> <p data-svelte-h="svelte-1frme0">Exploraremos los conceptos centrales de GRPO tal como aparecen en <code>GRPOTrainer</code> de TRL, usando fragmentos de la documentación oficial de TRL como guía.</p> <blockquote class="tip" data-svelte-h="svelte-11norxp"><p>Este capítulo está orientado a quienes empiezan con TRL. Si ya tienes experiencia, quizá también te interese revisar la <a href="https://github.com/huggingface/open-r1/blob/main/src/open_r1/grpo.py" rel="nofollow">implementación de Open R1</a> de GRPO.</p></blockquote> <p data-svelte-h="svelte-1ajhojj">Primero, recordemos algunos conceptos importantes del algoritmo GRPO:</p> <ul data-svelte-h="svelte-12kd6qs"><li>Formación de grupos: el modelo genera múltiples completions para cada prompt.</li> <li>Aprendizaje por preferencias: el modelo aprende a partir de una función de recompensa que compara grupos de completions.</li> <li>Configuración de entrenamiento: el modelo usa una configuración para controlar el proceso de entrenamiento.</li></ul> <p data-svelte-h="svelte-1dqejmx">¿Qué necesitamos para implementarlo?</p> <ul data-svelte-h="svelte-9mzu1v"><li>Definir un dataset de prompts.</li> <li>Definir una función de recompensa que reciba una lista de completions y devuelva una lista de recompensas.</li> <li>Configurar el entrenamiento con <code>GRPOConfig</code>.</li> <li>Entrenar el modelo con <code>GRPOTrainer</code>.</li></ul> <p data-svelte-h="svelte-uessia">Aquí tienes un ejemplo mínimo:</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="componentes-clave" 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="#componentes-clave"><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>Componentes clave</span></h2> <h3 class="relative group"><a id="1-formato-del-dataset" 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-formato-del-dataset"><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. Formato del dataset</span></h3> <p data-svelte-h="svelte-149cq7e">Tu dataset debe contener prompts a los que el modelo responderá. <code>GRPOTrainer</code> generará múltiples completions para cada prompt y usará la función de recompensa para compararlas.</p> <h3 class="relative group"><a id="2-función-de-recompensa" 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-función-de-recompensa"><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. Función de recompensa</span></h3> <p data-svelte-h="svelte-1mr8kty">La función de recompensa es crucial: determina cómo aprende el modelo. Aquí tienes dos ejemplos prácticos:</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-configuración-del-entrenamiento" 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-configuración-del-entrenamiento"><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. Configuración del entrenamiento</span></h3> <p data-svelte-h="svelte-8gixmy">Parámetros clave a considerar en <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(
	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>,
	per_device_train_batch_size=<span class="hljs-number">4</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>,
	use_vllm=<span class="hljs-literal">True</span>,
	)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1y5kihb">El parámetro <code>num_generation</code> es especialmente importante porque define el tamaño del grupo.</p> <ul data-svelte-h="svelte-4btybj"><li>Demasiado pequeño, por ejemplo 2 o 3, puede no aportar suficiente diversidad.</li> <li>Recomendado, entre 4 y 16, suele dar un buen equilibrio entre diversidad y coste.</li> <li>Valores más altos pueden mejorar el aprendizaje, pero incrementan bastante el coste computacional.</li></ul> <h2 class="relative group"><a id="consejos-para-tener-éxito" 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="#consejos-para-tener-éxito"><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>Consejos para tener éxito</span></h2> <ol data-svelte-h="svelte-17xjn5k"><li><strong>Gestión de memoria</strong>: ajusta <code>per_device_train_batch_size</code> y <code>gradient_accumulation_steps</code> según la memoria de tu GPU.</li> <li><strong>Velocidad</strong>: activa <code>use_vllm=True</code> para acelerar la generación si tu modelo es compatible.</li> <li><strong>Monitorización</strong>: observa métricas como <code>reward</code>, <code>reward_std</code> y <code>kl</code> durante el entrenamiento.</li></ol> <h2 class="relative group"><a id="diseño-de-funciones-de-recompensa" 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="#diseño-de-funciones-de-recompensa"><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>Diseño de funciones de recompensa</span></h2> <p data-svelte-h="svelte-vmayqv">El paper de DeepSeek R1 muestra varios enfoques eficaces que puedes adaptar.</p> <h3 class="relative group"><a id="1-recompensas-basadas-en-longitud" 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-recompensas-basadas-en-longitud"><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. Recompensas basadas en longitud</span></h3> <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-6dzchn">Esta función penaliza completions demasiado cortas o demasiado largas y empuja al modelo hacia una longitud objetivo.</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-recompensas-basadas-en-reglas-para-tareas-verificables" 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-recompensas-basadas-en-reglas-para-tareas-verificables"><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. Recompensas basadas en reglas para tareas verificables</span></h2> <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-keyword">try</span>:
	answer = extract_final_answer(completion)
	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>:
	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-recompensas-basadas-en-formato" 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-recompensas-basadas-en-formato"><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. Recompensas basadas en formato</span></h2> <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>
	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>:
	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-keyword">else</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_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-12j9xcv">Estos ejemplos muestran cómo puedes implementar funciones de recompensa inspiradas en el proceso de entrenamiento de DeepSeek R1.</p> <h2 class="relative group"><a id="eso-es-todo" 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="#eso-es-todo"><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>¡Eso es todo!</span></h2> <p data-svelte-h="svelte-249vsf">En la siguiente sección seguirás un ejercicio para implementar GRPO en 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/es/chapter12/4.mdx" 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_1nznq34 = {
	assets: "/docs/course/pr_1213/es",
	base: "/docs/course/pr_1213/es",
	env: {}
	};

	const element = document.currentScript.parentElement;

	const data = [null,null];

	Promise.all([
	import("/docs/course/pr_1213/es/_app/immutable/entry/start.36d27295.js"),
	import("/docs/course/pr_1213/es/_app/immutable/entry/app.3b43d7f3.js")
	]).then(([kit, app]) => {
	kit.start(app, element, {
	node_ids: [0, 32],
	data,
	form: null,
	error: null
	});
	});
	}
	</script>

Xet Storage Details

Size:: 39 kB
Xet hash:: bd43ec4a956aeac83e7695bb58b5193c17fe780acd71eb37e676c0a4ad85be5d

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