Hugging Face's logo

Buckets:
hf-doc-build
/
doc-dev

Files
xet
hf-doc-build/doc-dev / trl /pr_5313 /en /openreward.html
HuggingFaceDocBuilder's picture
HuggingFaceDocBuilder
download
raw
70.4 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;OpenReward Integration for Training LLMs with Environments&quot;,&quot;local&quot;:&quot;openreward-integration-for-training-llms-with-environments&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;When to use OpenReward environments&quot;,&quot;local&quot;:&quot;when-to-use-openreward-environments&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Installation&quot;,&quot;local&quot;:&quot;installation&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Quick start&quot;,&quot;local&quot;:&quot;quick-start&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Using a hub environment&quot;,&quot;local&quot;:&quot;using-a-hub-environment&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Using a self-hosted environment&quot;,&quot;local&quot;:&quot;using-a-self-hosted-environment&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Running a minimal environment locally&quot;,&quot;local&quot;:&quot;running-a-minimal-environment-locally&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Selecting tasks&quot;,&quot;local&quot;:&quot;selecting-tasks&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;How tool binding works&quot;,&quot;local&quot;:&quot;how-tool-binding-works&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Reward functions&quot;,&quot;local&quot;:&quot;reward-functions&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;OpenRewardSpec&quot;,&quot;local&quot;:&quot;trl.experimental.openreward.OpenRewardSpec&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Limitations&quot;,&quot;local&quot;:&quot;limitations&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Reference&quot;,&quot;local&quot;:&quot;reference&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}"/>
 <link href="/docs/trl/pr_5313/en/_app/immutable/entry/start.DU7fVq-0.js" rel="modulepreload">
 <link href="/docs/trl/pr_5313/en/_app/immutable/chunks/IyodssV2.js" rel="modulepreload">
 <link href="/docs/trl/pr_5313/en/_app/immutable/chunks/BxrqFSCq.js" rel="modulepreload">
 <link href="/docs/trl/pr_5313/en/_app/immutable/entry/app.2kk7sehm.js" rel="modulepreload">
 <link href="/docs/trl/pr_5313/en/_app/immutable/chunks/BxHPRFe5.js" rel="modulepreload">
 <link href="/docs/trl/pr_5313/en/_app/immutable/chunks/Df-tOpjM.js" rel="modulepreload">
 <link href="/docs/trl/pr_5313/en/_app/immutable/chunks/DsnmJJEf.js" rel="modulepreload">
 <link href="/docs/trl/pr_5313/en/_app/immutable/chunks/f6ZfvXQ7.js" rel="modulepreload">
 <link href="/docs/trl/pr_5313/en/_app/immutable/nodes/0.D4Lhtc7p.js" rel="modulepreload">
 <link href="/docs/trl/pr_5313/en/_app/immutable/chunks/CVE88MXi.js" rel="modulepreload">
 <link href="/docs/trl/pr_5313/en/_app/immutable/nodes/24.B-fcKN9Q.js" rel="modulepreload">
 <link href="/docs/trl/pr_5313/en/_app/immutable/chunks/ChUdhzCR.js" rel="modulepreload">
 <!--7mf4vw--><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;OpenReward Integration for Training LLMs with Environments&quot;,&quot;local&quot;:&quot;openreward-integration-for-training-llms-with-environments&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;When to use OpenReward environments&quot;,&quot;local&quot;:&quot;when-to-use-openreward-environments&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Installation&quot;,&quot;local&quot;:&quot;installation&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Quick start&quot;,&quot;local&quot;:&quot;quick-start&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Using a hub environment&quot;,&quot;local&quot;:&quot;using-a-hub-environment&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Using a self-hosted environment&quot;,&quot;local&quot;:&quot;using-a-self-hosted-environment&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Running a minimal environment locally&quot;,&quot;local&quot;:&quot;running-a-minimal-environment-locally&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Selecting tasks&quot;,&quot;local&quot;:&quot;selecting-tasks&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;How tool binding works&quot;,&quot;local&quot;:&quot;how-tool-binding-works&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Reward functions&quot;,&quot;local&quot;:&quot;reward-functions&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;OpenRewardSpec&quot;,&quot;local&quot;:&quot;trl.experimental.openreward.OpenRewardSpec&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Limitations&quot;,&quot;local&quot;:&quot;limitations&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Reference&quot;,&quot;local&quot;:&quot;reference&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}"/><!---->
 <link href="/docs/trl/pr_5313/en/_app/immutable/assets/0.tn0RQdqM.css" rel="modulepreload"> <!--[--><!--[0--><!--[--><!--[0--><!--[--><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> <!--[-1--><!--]--></div><!----> <!--[0--><h1 class="relative group"><a id="openreward-integration-for-training-llms-with-environments" 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="#openreward-integration-for-training-llms-with-environments"><span><svg 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>OpenReward Integration for Training LLMs with Environments</span></h1><!--]--><!----> <p><a href="https://openreward.ai" rel="nofollow">OpenReward</a> is an open ecosystem for RL environments built on the <a href="https://openrewardstandard.io" rel="nofollow">Open Reward Standard (ORS)</a> — a public, language-agnostic HTTP/SSE protocol for how an environment exposes its tasks, tools, sessions, and rewards. Because ORS is just a protocol, the same environment can run on the <a href="https://openreward.ai" rel="nofollow">OpenReward platform</a>, self-hosted on any container service, or locally on <code>localhost</code> for development. A catalog of ready-to-use environments is available at <a href="https://openreward.ai" rel="nofollow">openreward.ai</a>.</p> <p>This guide covers <strong>how to integrate OpenReward with TRL</strong>. For more on the standard itself, see the <a href="https://docs.openreward.ai/" rel="nofollow">ORS docs</a>.</p> <blockquote class="note"><p>The integration lives at <code>trl.experimental.openreward</code> and is gated behind the <code>trl[openreward]</code> extra (lazy-imported — non-users pay nothing).</p></blockquote> <!--[1--><h2 class="relative group"><a id="when-to-use-openreward-environments" 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="#when-to-use-openreward-environments"><span><svg 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>When to use OpenReward environments</span></h2><!--]--><!----> <p><a href="/docs/trl/pr_5313/en/gspo_token#trl.GRPOTrainer">GRPOTrainer</a> supports environment-based training via the <code>environment_factory</code> slot — see <a href="openenv">OpenEnv</a> for the general contract. Use OpenReward when you want to train against an ORS-speaking environment: the <a href="https://openreward.ai" rel="nofollow">OpenReward catalog</a> (e.g. <code>Eigent/SETA</code>, <code>kanishk/EndlessTerminals</code>, <code>nebius/SWE-rebench-V2</code>), an env you self-host on your own infra, or a local server you’re developing.</p> <!--[1--><h2 class="relative group"><a id="installation" 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="#installation"><span><svg 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>Installation</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 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="language-bash "><!---->pip install trl[openreward]<!----></pre></div><!----> <p>This installs the <code>openreward</code> Python SDK. The integration itself imports <code>openreward</code> lazily, so users who don’t touch <code>trl.experimental.openreward</code> aren’t affected.</p> <!--[1--><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 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>The <code>OpenRewardSpec</code> class wires a single ORS environment into the three TRL trainer slots — <code>train_dataset</code>, <code>environment_factory</code>, <code>reward_funcs</code> — by exposing properties that map 1:1 to those kwarg names:</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 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="language-python "><!----><span class="hljs-keyword">from</span> trl <span class="hljs-keyword">import</span> GRPOConfig, GRPOTrainer
<span class="hljs-keyword">from</span> trl.experimental.openreward <span class="hljs-keyword">import</span> OpenRewardSpec
spec = OpenRewardSpec(<span class="hljs-string">&quot;Eigent/SETA&quot;</span>, num_tasks=<span class="hljs-number">64</span>)
trainer = GRPOTrainer(
 model=<span class="hljs-string">&quot;Qwen/Qwen3-4B&quot;</span>,
 args=GRPOConfig(
 num_generations=<span class="hljs-number">2</span>,
 max_steps=<span class="hljs-number">5</span>,
 max_tool_calling_iterations=<span class="hljs-number">20</span>,
 log_completions=<span class="hljs-literal">True</span>,
 ),
 train_dataset=spec.train_dataset,
 environment_factory=spec.environment_factory,
 reward_funcs=spec.reward_funcs,
)
trainer.train()<!----></pre></div><!----> <p>Under the hood <code>OpenRewardSpec</code> does three things, lazily on first access:</p> <ol><li><strong><code>spec.train_dataset</code></strong>: derives a <code>datasets.Dataset</code> from the env’s task list (one HTTP roundtrip via the SDK). Has at minimum <code>prompt</code>, <code>task_index</code>, plus per-task metadata columns folded in.</li> <li><strong><code>spec.environment_factory</code></strong>: returns a zero-arg callable that produces a fresh per-rollout adapter on each call. The adapter exposes one Python method per ORS tool, with a typed signature and docstring auto-generated from the env’s JSON Schema. TRL’s tool collector picks them up via <code>inspect.getmembers</code>.</li> <li><strong><code>spec.reward_funcs</code></strong>: an outcome-only reward function (last non-null reward in the trajectory) suitable for sparse-reward envs like SETA.</li></ol> <!--[1--><h2 class="relative group"><a id="using-a-hub-environment" 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-hub-environment"><span><svg 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 hub environment</span></h2><!--]--><!----> <p>Pass an <a href="https://openreward.ai" rel="nofollow">openreward.ai</a> catalog name as the target. The SDK reads <code>OPENREWARD_API_KEY</code> from the environment for authentication.</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 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="language-python "><!---->spec = OpenRewardSpec(<span class="hljs-string">&quot;Eigent/SETA&quot;</span>, num_tasks=<span class="hljs-number">64</span>)<!----></pre></div><!----> <!--[1--><h2 class="relative group"><a id="using-a-self-hosted-environment" 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-self-hosted-environment"><span><svg 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 self-hosted environment</span></h2><!--]--><!----> <p>Pass the URL directly. No API key is needed if your server doesn’t enforce one.</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 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="language-python "><!---->spec = OpenRewardSpec(<span class="hljs-string">&quot;https://my-org-my-env.hf.space&quot;</span>, env_name=<span class="hljs-string">&quot;my_env&quot;</span>)<!----></pre></div><!----> <blockquote class="important"><p>The <code>openreward</code> SDK by default expects a two-subdomain platform layout (<code>api.&lt;host></code> for stateless calls and <code>sessions.&lt;host></code> for SSE-based session calls). For <strong>single-host</strong> self-hosted servers (one URL serving everything), set the override env vars below before constructing <code>OpenRewardSpec</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 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="language-python "><!----><span class="hljs-keyword">import</span> os
URL = <span class="hljs-string">&quot;https://my-org-my-env.hf.space&quot;</span>
os.environ[<span class="hljs-string">&quot;OPENREWARD_API_URL&quot;</span>] = URL
os.environ[<span class="hljs-string">&quot;OPENREWARD_SESSION_URL&quot;</span>] = URL
spec = OpenRewardSpec(URL, env_name=<span class="hljs-string">&quot;my_env&quot;</span>)<!----></pre></div><!----></blockquote> <!--[1--><h2 class="relative group"><a id="running-a-minimal-environment-locally" 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="#running-a-minimal-environment-locally"><span><svg 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>Running a minimal environment locally</span></h2><!--]--><!----> <p>The fastest way to try the integration end-to-end without external dependencies is a tiny ORS server defined with the <code>openreward</code> SDK’s <code>Environment</code> + <code>Server</code> scaffolding. The example below is a complete <code>echo</code> environment — the model wins by calling <code>echo(text=...)</code> with the task’s target string.</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 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="language-python "><!----><span class="hljs-comment"># server.py</span>
<span class="hljs-keyword">from</span> pydantic <span class="hljs-keyword">import</span> BaseModel
<span class="hljs-keyword">from</span> openreward.environments <span class="hljs-keyword">import</span> Environment, JSONObject, Server, TextBlock, ToolOutput, tool
<span class="hljs-keyword">class</span> <span class="hljs-title class_">EchoTaskSpec</span>(<span class="hljs-title class_ inherited__">BaseModel</span>):
 target: <span class="hljs-built_in">str</span>
<span class="hljs-keyword">class</span> <span class="hljs-title class_">EchoParams</span>(<span class="hljs-title class_ inherited__">BaseModel</span>):
 text: <span class="hljs-built_in">str</span>
<span class="hljs-keyword">class</span> <span class="hljs-title class_">EchoEnvironment</span>(<span class="hljs-title class_ inherited__">Environment</span>):
 <span class="hljs-keyword">def</span> <span class="hljs-title function_">__init__</span>(<span class="hljs-params">self, task_spec: JSONObject = {}, secrets: <span class="hljs-built_in">dict</span>[<span class="hljs-built_in">str</span>, <span class="hljs-built_in">str</span>] = {}</span>):
 <span class="hljs-built_in">super</span>().__init__(task_spec)
 <span class="hljs-variable language_">self</span>.config = EchoTaskSpec.model_validate(task_spec)
<span class="hljs-meta"> @classmethod</span>
 <span class="hljs-keyword">def</span> <span class="hljs-title function_">list_splits</span>(<span class="hljs-params">cls</span>) -&gt; <span class="hljs-built_in">list</span>[<span class="hljs-built_in">str</span>]:
 <span class="hljs-keyword">return</span> [<span class="hljs-string">&quot;train&quot;</span>]
<span class="hljs-meta"> @classmethod</span>
 <span class="hljs-keyword">def</span> <span class="hljs-title function_">list_tasks</span>(<span class="hljs-params">cls, split: <span class="hljs-built_in">str</span></span>) -&gt; <span class="hljs-built_in">list</span>[JSONObject]:
 <span class="hljs-keyword">return</span> [{<span class="hljs-string">&quot;target&quot;</span>: <span class="hljs-string">&quot;hello&quot;</span>}, {<span class="hljs-string">&quot;target&quot;</span>: <span class="hljs-string">&quot;world&quot;</span>}]
<span class="hljs-keyword">def</span> <span class="hljs-title function_">get_prompt</span>(<span class="hljs-params">self</span>) -&gt; <span class="hljs-built_in">list</span>[TextBlock]:
 <span class="hljs-keyword">return</span> [TextBlock(<span class="hljs-built_in">type</span>=<span class="hljs-string">&quot;text&quot;</span>, text=<span class="hljs-string">f&quot;Echo &#x27;<span class="hljs-subst">{self.config.target}</span>&#x27; to win.&quot;</span>)]
<span class="hljs-meta"> @tool</span>
 <span class="hljs-keyword">async</span> <span class="hljs-keyword">def</span> <span class="hljs-title function_">echo</span>(<span class="hljs-params">self, params: EchoParams</span>) -&gt; ToolOutput:
 <span class="hljs-string">&quot;&quot;&quot;Submit a string. Reward 1.0 + finished if it matches the target.
Args:
 text: The string to echo back.
 &quot;&quot;&quot;</span>
 correct = params.text == <span class="hljs-variable language_">self</span>.config.target
 <span class="hljs-keyword">return</span> ToolOutput(
 blocks=[TextBlock(<span class="hljs-built_in">type</span>=<span class="hljs-string">&quot;text&quot;</span>, text=<span class="hljs-string">&quot;match&quot;</span> <span class="hljs-keyword">if</span> correct <span class="hljs-keyword">else</span> <span class="hljs-string">&quot;no match&quot;</span>)],
 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">0.0</span>,
 finished=correct,
 )
<span class="hljs-keyword">if</span> __name__ == <span class="hljs-string">&quot;__main__&quot;</span>:
 Server([EchoEnvironment]).run(host=<span class="hljs-string">&quot;0.0.0.0&quot;</span>, port=<span class="hljs-number">8000</span>)<!----></pre></div><!----> <p>Run it:</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 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="language-bash "><!---->pip install openreward fastapi uvicorn pydantic
python server.py <span class="hljs-comment"># listens on :8000</span><!----></pre></div><!----> <p>Then point <code>OpenRewardSpec</code> at it (with the URL overrides described above):</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 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="language-python "><!----><span class="hljs-keyword">import</span> os
URL = <span class="hljs-string">&quot;http://127.0.0.1:8000&quot;</span>
os.environ[<span class="hljs-string">&quot;OPENREWARD_API_URL&quot;</span>] = URL
os.environ[<span class="hljs-string">&quot;OPENREWARD_SESSION_URL&quot;</span>] = URL
<span class="hljs-keyword">from</span> trl.experimental.openreward <span class="hljs-keyword">import</span> OpenRewardSpec
spec = OpenRewardSpec(URL, env_name=<span class="hljs-string">&quot;echoenvironment&quot;</span>)
<span class="hljs-built_in">print</span>(spec.train_dataset) <span class="hljs-comment"># 2 rows, task_index + target columns</span><!----></pre></div><!----> <p>This is also the fixture pattern used by TRL’s own tests — see <a href="https://huggingface.co/spaces/trl-internal-testing/openreward-echo-env" rel="nofollow"><code>trl-internal-testing/openreward-echo-env</code></a> for the deployed Space.</p> <!--[1--><h2 class="relative group"><a id="selecting-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="#selecting-tasks"><span><svg 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>Selecting tasks</span></h2><!--]--><!----> <p><code>OpenRewardSpec</code> accepts either a count or an explicit index 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 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="language-python "><!---->spec = OpenRewardSpec(<span class="hljs-string">&quot;Eigent/SETA&quot;</span>, num_tasks=<span class="hljs-number">10</span>) <span class="hljs-comment"># first 10 tasks</span>
spec = OpenRewardSpec(<span class="hljs-string">&quot;Eigent/SETA&quot;</span>, indices=[<span class="hljs-number">0</span>, <span class="hljs-number">5</span>, <span class="hljs-number">13</span>, <span class="hljs-number">27</span>]) <span class="hljs-comment"># specific indices</span>
spec = OpenRewardSpec(<span class="hljs-string">&quot;Eigent/SETA&quot;</span>, indices=<span class="hljs-built_in">list</span>(<span class="hljs-built_in">range</span>(<span class="hljs-number">50</span>, <span class="hljs-number">100</span>))) <span class="hljs-comment"># range</span><!----></pre></div><!----> <p><code>num_tasks</code> and <code>indices</code> are mutually exclusive and both fetch only the tasks they need (no full task list scan).</p> <!--[1--><h2 class="relative group"><a id="how-tool-binding-works" 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="#how-tool-binding-works"><span><svg 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>How tool binding works</span></h2><!--]--><!----> <p>At construction the spec calls the env’s <code>/tools</code> endpoint to fetch a list of tool specs (each with a name, description, and JSON Schema for arguments). For each tool it generates a Python method on the per-rollout adapter with a typed signature and a docstring derived from the schema. So <code>transformers.utils.get_json_schema</code> and TRL’s <code>inspect.getmembers(env, ismethod)</code> both produce the right tool schema for the model with no per-env wrapper code.</p> <p>If a tool description contains characters that aren’t safe to splice into Python source, the binder falls back to a sanitized form so binding never fails on real envs.</p> <!--[1--><h2 class="relative group"><a id="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="#reward-functions"><span><svg 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 functions</span></h2><!--]--><!----> <p><code>spec.reward_funcs</code> defaults to an outcome-only reward — for each rollout it returns the last non-null reward observed during the trajectory. This is the right default for sparse-reward envs (e.g. SETA, where only <code>submit_solution</code> returns a non-null reward).</p> <p>If you want a custom reward, write a regular TRL reward function and pass it directly:</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 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="language-python "><!----><span class="hljs-keyword">def</span> <span class="hljs-title function_">my_reward</span>(<span class="hljs-params">environments, **kwargs</span>) -&gt; <span class="hljs-built_in">list</span>[<span class="hljs-built_in">float</span>]:
 <span class="hljs-keyword">return</span> [env.reward * <span class="hljs-number">2.0</span> <span class="hljs-keyword">for</span> env <span class="hljs-keyword">in</span> environments] <span class="hljs-comment"># double the env reward, etc.</span>
trainer = GRPOTrainer(
 ...,
 reward_funcs=my_reward,
)<!----></pre></div><!----> <p>The per-rollout adapter exposes the running state TRL needs — <code>env.reward</code>, <code>env.rewards</code>, <code>env.metadata</code>, <code>env.finished</code>, <code>env.last_output</code> — for arbitrary post-hoc reward shaping.</p> <!--[1--><h2 class="relative group"><a id="trl.experimental.openreward.OpenRewardSpec" 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.experimental.openreward.OpenRewardSpec"><span><svg 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>OpenRewardSpec</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.experimental.openreward.OpenRewardSpec"><!----><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.experimental.openreward.</span><span class="font-semibold">OpenRewardSpec</span></span></h3><!----> <a id="trl.experimental.openreward.OpenRewardSpec" class="header-link invisible with-hover:group-hover:visible pr-2" href="#trl.experimental.openreward.OpenRewardSpec"><svg 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> <!--[0--><a class="!ml-auto !text-gray-400 !no-underline text-sm flex items-center" href="https://github.com/huggingface/trl/blob/vr_5313/trl/experimental/openreward/_spec.py#L78" target="_blank"><span>&lt;</span> <span class="hidden md:block mx-0.5 hover:!underline">source</span> <span>></span></a><!--]--></span> <!--[0--><p class="font-mono text-xs md:text-sm !leading-relaxed !my-6"><span>(</span> <!--[--><span class="comma cursor-pointer"><span class="rounded hover:bg-black hover:text-white dark:hover:bg-white dark:hover:text-black">target<span class="opacity-60">: str</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">num_tasks<span class="opacity-60">: int | 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">split<span class="opacity-60">: str = 'train'</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">indices<span class="opacity-60">: list[int] | 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">api_key<span class="opacity-60">: str | 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">secrets<span class="opacity-60">: dict[str, str] | 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">env_name<span class="opacity-60">: str | 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">include_metadata<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">discover_task_tools<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">task_tools_discovery_index<span class="opacity-60">: int | None = None</span></span></span><!--]--> <span>)</span> <!--[-1--><!--]--></p><!--]--> <div class="!mb-10 relative docstring-details "><!--[-1--><!--]--> <!--[0--><p class="flex items-center font-semibold !mt-2 !mb-2 text-gray-800">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.experimental.openreward.OpenRewardSpec.target" 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.experimental.openreward.OpenRewardSpec.target"><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><!----><strong>target</strong> (<em>str</em>) &#x2014;
Either an openreward.ai catalog name (<em>&#x201C;Eigent/SETA&#x201D;</em>) or a URL pointing at any ORS server
(<em>&#x201D;<a href="https://you-seta.hf.space%22" rel="nofollow">https://you-seta.hf.space&#x201D;</a></em>, <em>&#x201C;http://localhost:8080&#x201D;</em>). Auto-detected by the presence of <em>://</em> in the
string.<!----></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.experimental.openreward.OpenRewardSpec.num_tasks" 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.experimental.openreward.OpenRewardSpec.num_tasks"><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><!----><strong>num_tasks</strong> (<em>int</em>, <em>optional</em>) &#x2014;
Cap on the number of tasks pulled into the dataset. <code>None</code> uses every task the env exposes.<!----></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.experimental.openreward.OpenRewardSpec.split" 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.experimental.openreward.OpenRewardSpec.split"><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><!----><strong>split</strong> (<em>str</em>, <em>optional</em>, defaults to <em>&#x201C;train&#x201D;</em>) &#x2014;
Which split&#x2019;s task list to draw from.<!----></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.experimental.openreward.OpenRewardSpec.indices" 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.experimental.openreward.OpenRewardSpec.indices"><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><!----><strong>indices</strong> (<em>list[int]</em>, <em>optional</em>) &#x2014;
Specific task indices to train on. Mutually exclusive with <code>num_tasks</code>. Useful for debugging or
curriculum subsets.<!----></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.experimental.openreward.OpenRewardSpec.api_key" 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.experimental.openreward.OpenRewardSpec.api_key"><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><!----><strong>api_key</strong> (<em>str</em>, <em>optional</em>) &#x2014;
<code>OPENREWARD_API_KEY</code> override. Only used when <code>target</code> is a catalog name.<!----></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.experimental.openreward.OpenRewardSpec.secrets" 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.experimental.openreward.OpenRewardSpec.secrets"><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><!----><strong>secrets</strong> (<em>dict[str, str]</em>, <em>optional</em>) &#x2014;
Per-session secrets forwarded to <code>env.session(secrets=)</code>.<!----></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.experimental.openreward.OpenRewardSpec.env_name" 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.experimental.openreward.OpenRewardSpec.env_name"><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><!----><strong>env_name</strong> (<em>str</em>, <em>optional</em>) &#x2014;
Override for the env name to look up on the server. Rarely needed.<!----></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.experimental.openreward.OpenRewardSpec.include_metadata" 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.experimental.openreward.OpenRewardSpec.include_metadata"><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><!----><strong>include_metadata</strong> (<em>bool</em>, <em>optional</em>, defaults to <em>True</em>) &#x2014;
Fold per-task metadata (<em>difficulty</em>, <em>category</em>, <em>tags</em>, &#x2026;) into the dataset rows so reward funcs can
read them via TRL&#x2019;s <code>inputs</code> argument.<!----></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.experimental.openreward.OpenRewardSpec.discover_task_tools" 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.experimental.openreward.OpenRewardSpec.discover_task_tools"><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><!----><strong>discover_task_tools</strong> (<em>bool</em>, <em>optional</em>, defaults to <em>True</em>) &#x2014;
If <code>True</code>, opens a short-lived ORS session and uses <code>session.list_tools()</code> so task-specific tools
(<code>GET &#x2026;/task_tools</code> per ORS &#x2014; e.g. <code>@tool(shared=False)</code> and <code>list_task_tools()</code>) are bound for GRPO.
If probe fails, falls back to <code>environment.list_tools()</code> only. Set <code>False</code> to skip the extra session
(shared tools only / offline quirks).<!----></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.experimental.openreward.OpenRewardSpec.task_tools_discovery_index" 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.experimental.openreward.OpenRewardSpec.task_tools_discovery_index"><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><!----><strong>task_tools_discovery_index</strong> (<em>int</em>, <em>optional</em>) &#x2014;
Task index used only for the discovery session when set; <strong>overrides</strong> multi-index probing below. When
omitted and <code>indices=</code> is set, discovery opens one probe session per <strong>distinct</strong> entry in <code>indices</code>
(sorted) and <strong>merges</strong> tool specs by name so task-specific tools from every listed task are bound. When
omitted and <code>num_tasks</code> / full-list mode is used, probes task <code>0</code> only. Ignored when
<code>discover_task_tools=False</code>.<!----></span></span></li><!--]--></ul><!--]--> <!--[-1--><!--]--> <!--[-1--><!--]--> <!--[-1--><!--]--></div></div><!----> <p>Single spec object that wires an ORS environment into a TRL trainer.</p></div> <!--[1--><h2 class="relative group"><a id="limitations" 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="#limitations"><span><svg 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>Limitations</span></h2><!--]--><!----> <ul><li>The integration is in <code>trl.experimental</code> — APIs may change. Set <code>TRL_EXPERIMENTAL_SILENCE=1</code> to silence the warning in CI logs.</li> <li>Currently exposes a single <code>OpenRewardSpec</code> covering one environment; multi-environment training (à la the OpenEnv “meta-environment” pattern) is not supported yet.</li> <li>Long-running rollouts (>15 min per episode) need a keepalive ping — not yet wired.</li></ul> <!--[1--><h2 class="relative group"><a id="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="#reference"><span><svg 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>Reference</span></h2><!--]--><!----> <ul><li><a href="https://openrewardstandard.io" rel="nofollow">Open Reward Standard</a></li> <li><a href="https://openreward.ai" rel="nofollow">OpenReward platform</a></li> <li><a href="https://pypi.org/project/openreward/" rel="nofollow"><code>openreward</code> Python SDK</a></li> <li><a href="https://huggingface.co/spaces/trl-internal-testing/openreward-echo-env" rel="nofollow">Echo env Space — <code>trl-internal-testing/openreward-echo-env</code></a></li></ul> <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/openreward.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><span class="underline">Update</span> on GitHub</span></a><!----> <p></p><!--]--><!----><!--]--><!--]--><!--]--> <!--[-1--><!--]--><!--]-->
<script>
 {
 __sveltekit_1w3vpa5 = {
 base: "/docs/trl/pr_5313/en",
 assets: "/docs/trl/pr_5313/en"
 };
 const element = document.currentScript.parentElement;
 Promise.all([
 import("/docs/trl/pr_5313/en/_app/immutable/entry/start.DU7fVq-0.js"),
 import("/docs/trl/pr_5313/en/_app/immutable/entry/app.2kk7sehm.js")
 ]).then(([kit, app]) => {
 kit.start(app, element, {
 node_ids: [0, 24],
 data: [null,null],
 form: null,
 error: null
 });
 });
 }
 </script>

Xet Storage Details

Size:
70.4 kB
·
Xet hash:
772f0b39304594181c53f9d51ab97fc9f48bb3bd28ffb95b96e387fe64a2987f

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