{"id":51114,"date":"2025-08-28T03:34:39","date_gmt":"2025-08-27T19:34:39","guid":{"rendered":"https:\/\/www.kdjingpai.com\/?p=51114"},"modified":"2025-08-28T03:34:39","modified_gmt":"2025-08-27T19:34:39","slug":"verifiers","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/en\/verifiers\/","title":{"rendered":"Verifiers\uff1a\u7528\u4e8e\u8bad\u7ec3\u5927\u8bed\u8a00\u6a21\u578b\u7684\u5f3a\u5316\u5b66\u4e60\u73af\u5883\u5de5\u5177\u5e93"},"content":{"rendered":"<p>Verifiers \u662f\u4e00\u4e2a\u7528\u4e8e\u521b\u5efa\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u73af\u5883\u548c\u8bad\u7ec3\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u4ee3\u7406\u7684\u6a21\u5757\u5316\u7ec4\u4ef6\u5e93\u3002 \u8fd9\u4e2a\u9879\u76ee\u7684\u76ee\u6807\u662f\u63d0\u4f9b\u4e00\u5957\u53ef\u9760\u7684\u5de5\u5177\uff0c\u8ba9\u5f00\u53d1\u8005\u53ef\u4ee5\u65b9\u4fbf\u5730\u6784\u5efa\u3001\u8bad\u7ec3\u548c\u8bc4\u4f30LLM\u4ee3\u7406\u3002 Verifiers \u5305\u542b\u4e00\u4e2a\u57fa\u4e8e\u00a0<code>transformers<\/code>\u00a0Trainer \u5b9e\u73b0\u7684\u5f02\u6b65 GRPO\uff08Generalized Reinforcement Learning with Policy Optimization\uff09\u8bad\u7ec3\u5668\uff0c\u5e76\u4e14\u5f97\u5230\u4e86\u00a0<code>prime-rl<\/code>\u00a0\u9879\u76ee\u7684\u652f\u6301\uff0c\u53ef\u7528\u4e8e\u5927\u89c4\u6a21 FSDP\uff08Fully Sharded Data Parallel\uff09\u8bad\u7ec3\u3002 \u9664\u4e86\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\uff0cVerifiers \u4e5f\u53ef\u4ee5\u76f4\u63a5\u7528\u4e8e\u6784\u5efa LLM \u8bc4\u4f30\u3001\u521b\u5efa\u5408\u6210\u6570\u636e\u7ba1\u9053\u548c\u5b9e\u73b0\u4ee3\u7406\u63a7\u5236\u7a0b\u5e8f\u3002 \u8be5\u9879\u76ee\u65e8\u5728\u6210\u4e3a\u4e00\u4e2a\u53ef\u9760\u7684\u5de5\u5177\u5305\uff0c\u5c3d\u91cf\u51cf\u5c11\u5728\u5f3a\u5316\u5b66\u4e60\u57fa\u7840\u8bbe\u65bd\u751f\u6001\u7cfb\u7edf\u4e2d\u5e38\u89c1\u7684\u201c\u4ee3\u7801\u5e93\u5206\u53c9\u6cdb\u6ee5\u201d\u95ee\u9898\uff0c\u4e3a\u5f00\u53d1\u8005\u63d0\u4f9b\u4e00\u4e2a\u7a33\u5b9a\u7684\u5f00\u53d1\u57fa\u7840\u3002<\/p>\n<h2>\u529f\u80fd\u5217\u8868<\/h2>\n<ul>\n<li><strong>\u6a21\u5757\u5316\u73af\u5883\u7ec4\u4ef6<\/strong>: \u63d0\u4f9b\u4e86\u4e00\u5957\u7528\u4e8e\u6784\u5efa\u5f3a\u5316\u5b66\u4e60\u73af\u5883\u7684\u6a21\u5757\u5316\u7ec4\u4ef6\uff0c\u4f7f\u73af\u5883\u7684\u521b\u5efa\u548c\u5b9a\u5236\u66f4\u52a0\u7b80\u5355\u3002<\/li>\n<li><strong>\u591a\u79cd\u73af\u5883\u7c7b\u578b\u652f\u6301<\/strong>:\n<ul>\n<li><code>SingleTurnEnv<\/code>: \u9002\u7528\u4e8e\u6bcf\u4e2a\u63d0\u793a\u53ea\u9700\u8981\u6a21\u578b\u5355\u6b21\u54cd\u5e94\u7684\u4efb\u52a1\u3002<\/li>\n<li><code>ToolEnv<\/code>: \u652f\u6301\u5229\u7528\u6a21\u578b\u539f\u751f\u7684\u5de5\u5177\u6216\u51fd\u6570\u8c03\u7528\u80fd\u529b\uff0c\u6784\u5efa\u4ee3\u7406\u5faa\u73af\u3002<\/li>\n<li><code>MultiTurnEnv<\/code>: \u4e3a\u7f16\u5199\u81ea\u5b9a\u4e49\u73af\u5883\u4ea4\u4e92\u534f\u8bae\u63d0\u4f9b\u4e86\u63a5\u53e3\uff0c\u9002\u7528\u4e8e\u591a\u8f6e\u5bf9\u8bdd\u6216\u4ea4\u4e92\u5f0f\u4efb\u52a1\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u5185\u7f6e\u8bad\u7ec3\u5668<\/strong>: \u5305\u542b\u4e00\u4e2a\u00a0<code>GRPOTrainer<\/code>\uff0c\u5b83\u4f7f\u7528\u00a0<code><a href=\"https:\/\/www.kdjingpai.com\/en\/vllm\/\">vLLM<\/a><\/code>\u00a0\u8fdb\u884c\u63a8\u7406\uff0c\u652f\u6301\u901a\u8fc7 Accelerate\/DeepSpeed \u8fd0\u884c <a href=\"https:\/\/www.kdjingpai.com\/en\/grpo-ruhezaishi\/\">GRPO<\/a> \u98ce\u683c\u7684\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u3002<\/li>\n<li><strong>\u547d\u4ee4\u884c\u5de5\u5177<\/strong>: \u63d0\u4f9b\u5b9e\u7528\u547d\u4ee4\u884c\u5de5\u5177\u7b80\u5316\u5de5\u4f5c\u6d41\u7a0b\uff1a\n<ul>\n<li><code>vf-init<\/code>: \u521d\u59cb\u5316\u4e00\u4e2a\u65b0\u7684\u73af\u5883\u6a21\u5757\u6a21\u677f\u3002<\/li>\n<li><code>vf-install<\/code>: \u5c06\u73af\u5883\u6a21\u5757\u5b89\u88c5\u5230\u5f53\u524d\u9879\u76ee\u4e2d\u3002<\/li>\n<li><code>vf-eval<\/code>: \u4f7f\u7528 API \u6a21\u578b\u5feb\u901f\u5bf9\u73af\u5883\u8fdb\u884c\u8bc4\u4f30\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u96c6\u6210\u4e0e\u517c\u5bb9\u6027<\/strong>: \u53ef\u4ee5\u8f7b\u677e\u96c6\u6210\u5230\u4efb\u4f55\u652f\u6301 OpenAI \u517c\u5bb9\u63a8\u7406\u5ba2\u6237\u7aef\u7684\u5f3a\u5316\u5b66\u4e60\u6846\u67b6\u4e2d\uff0c\u5e76\u539f\u751f\u652f\u6301\u4e0e\u00a0<code>prime-rl<\/code>\u00a0\u534f\u540c\u5de5\u4f5c\uff0c\u4ee5\u5b9e\u73b0\u66f4\u9ad8\u6548\u3001\u66f4\u5927\u89c4\u6a21\u7684\u8bad\u7ec3\u3002<\/li>\n<li><strong>\u7075\u6d3b\u7684\u5956\u52b1\u673a\u5236<\/strong>: \u901a\u8fc7\u00a0<code>Rubric<\/code>\u00a0\u7c7b\u5c01\u88c5\u4e00\u4e2a\u6216\u591a\u4e2a\u5956\u52b1\u51fd\u6570\uff0c\u53ef\u4ee5\u4e3a\u6a21\u578b\u751f\u6210\u7684\u5b8c\u6210\u60c5\u51b5\u5b9a\u4e49\u590d\u6742\u7684\u8bc4\u4f30\u6807\u51c6\u3002<\/li>\n<\/ul>\n<h2>\u4f7f\u7528\u5e2e\u52a9<\/h2>\n<p>Verifiers \u5e93\u5efa\u8bae\u4e0e\u00a0<code>uv<\/code>\u00a0\u5305\u7ba1\u7406\u5668\u4e00\u8d77\u5728\u4f60\u7684\u9879\u76ee\u4e2d\u4f7f\u7528\u3002<\/p>\n<h3>1. \u5b89\u88c5<\/h3>\n<p>\u9996\u5148\uff0c\u4f60\u9700\u8981\u521b\u5efa\u4e00\u4e2a\u65b0\u7684\u865a\u62df\u73af\u5883\u5e76\u6fc0\u6d3b\u5b83\u3002<\/p>\n<pre><code># \u5b89\u88c5 uv\r\ncurl -LsSf https:\/\/astral.sh\/uv\/install.sh | sh\r\n# \u521d\u59cb\u5316\u4e00\u4e2a\u65b0\u9879\u76ee\r\nuv init\r\n# \u6fc0\u6d3b\u865a\u62df\u73af\u5883\r\nsource .venv\/bin\/activate\r\n<\/code><\/pre>\n<p>\u63a5\u4e0b\u6765\uff0c\u6839\u636e\u4f60\u7684\u9700\u6c42\u5b89\u88c5 Verifiers\uff1a<\/p>\n<ul>\n<li><strong>\u672c\u5730\u5f00\u53d1\u4e0e\u8bc4\u4f30 (CPU)<\/strong>: \u5982\u679c\u4f60\u53ea\u4f7f\u7528 API \u6a21\u578b\u8fdb\u884c\u5f00\u53d1\u548c\u8bc4\u4f30\uff0c\u5b89\u88c5\u6838\u5fc3\u5e93\u5373\u53ef\u3002\n<pre><code># \u5b89\u88c5\u6838\u5fc3\u5e93\r\nuv add verifiers\r\n# \u5982\u679c\u9700\u8981 Jupyter \u548c\u6d4b\u8bd5\u652f\u6301\r\nuv add 'verifiers[dev]'\r\n<\/code><\/pre>\n<\/li>\n<li><strong>GPU \u8bad\u7ec3<\/strong>: \u5982\u679c\u4f60\u8ba1\u5212\u4f7f\u7528\u00a0<code>vf.GRPOTrainer<\/code>\u00a0\u5728 GPU \u4e0a\u8fdb\u884c\u6a21\u578b\u8bad\u7ec3\uff0c\u9700\u8981\u5b89\u88c5\u5305\u542b\u6240\u6709\u4f9d\u8d56\u7684\u7248\u672c\uff0c\u5e76\u989d\u5916\u5b89\u88c5\u00a0<code>flash-attn<\/code>\u3002\n<pre><code>uv add 'verifiers[all]' &amp;&amp; uv pip install flash-attn --no-build-isolation\r\n<\/code><\/pre>\n<\/li>\n<li><strong>\u4f7f\u7528\u6700\u65b0\u7684\u5f00\u53d1\u7248\u672c<\/strong>: \u4f60\u4e5f\u53ef\u4ee5\u76f4\u63a5\u4ece GitHub \u7684\u00a0<code>main<\/code>\u00a0\u5206\u652f\u5b89\u88c5\u3002\n<pre><code>uv add verifiers @ git+https:\/\/github.com\/willccbb\/verifiers.git\r\n<\/code><\/pre>\n<\/li>\n<li><strong>\u4ece\u6e90\u7801\u5b89\u88c5 (\u6838\u5fc3\u5e93\u5f00\u53d1)<\/strong>: \u5982\u679c\u4f60\u9700\u8981\u4fee\u6539 Verifiers \u6838\u5fc3\u5e93\uff0c\u53ef\u4ee5\u4ece\u6e90\u7801\u5b89\u88c5\u3002\n<pre><code>git clone https:\/\/github.com\/willccbb\/verifiers.git\r\ncd verifiers\r\nuv <a href=\"https:\/\/www.kdjingpai.com\/en\/sync\/\">sync<\/a> --all-extras &amp;&amp; uv pip install flash-attn --no-build-isolation\r\nuv run pre-commit install\r\n<\/code><\/pre>\n<\/li>\n<\/ul>\n<h3>2. \u521b\u5efa\u548c\u7ba1\u7406\u73af\u5883<\/h3>\n<p>Verifiers \u5c06\u6bcf\u4e2a\u5f3a\u5316\u5b66\u4e60\u73af\u5883\u89c6\u4e3a\u4e00\u4e2a\u53ef\u5b89\u88c5\u7684 Python \u6a21\u5757\u3002<\/p>\n<ul>\n<li><strong>\u521d\u59cb\u5316\u4e00\u4e2a\u65b0\u73af\u5883<\/strong>: \u4f7f\u7528\u00a0<code>vf-init<\/code>\u00a0\u547d\u4ee4\u521b\u5efa\u4e00\u4e2a\u65b0\u7684\u73af\u5883\u6a21\u677f\u3002\n<pre><code># \u521b\u5efa\u4e00\u4e2a\u540d\u4e3a my-new-env \u7684\u73af\u5883\r\nvf-init my-new-env\r\n<\/code><\/pre>\n<p>\u8fd9\u4e2a\u547d\u4ee4\u4f1a\u5728\u00a0<code>environments\/my-new-env<\/code>\u00a0\u76ee\u5f55\u4e0b\u751f\u6210\u4e00\u4e2a\u5305\u542b\u00a0<code>pyproject.toml<\/code>\u00a0\u548c\u57fa\u672c\u7ed3\u6784\u7684\u73af\u5883\u6a21\u677f\u3002<\/li>\n<li><strong>\u5b89\u88c5\u73af\u5883<\/strong>: \u521b\u5efa\u540e\uff0c\u4f7f\u7528\u00a0<code>vf-install<\/code>\u00a0\u5c06\u5176\u5b89\u88c5\u5230\u4f60\u7684 Python \u73af\u5883\u4e2d\uff0c\u4f7f\u5176\u53ef\u4ee5\u88ab\u5bfc\u5165\u548c\u4f7f\u7528\u3002\n<pre><code># \u5b89\u88c5\u672c\u5730\u73af\u5883\r\nvf-install my-new-env\r\n# \u4f60\u4e5f\u53ef\u4ee5\u76f4\u63a5\u4ece verifiers \u5b98\u65b9\u4ed3\u5e93\u5b89\u88c5\u793a\u4f8b\u73af\u5883\r\nvf-install vf-math-python --from-repo\r\n<\/code><\/pre>\n<\/li>\n<\/ul>\n<h3>3. \u4f7f\u7528\u73af\u5883<\/h3>\n<p>\u5b89\u88c5\u73af\u5883\u540e\uff0c\u4f60\u53ef\u4ee5\u4f7f\u7528\u00a0<code>vf.load_environment<\/code>\u00a0\u51fd\u6570\u52a0\u8f7d\u5b83\uff0c\u5e76\u8fdb\u884c\u8bc4\u4f30\u6216\u8bad\u7ec3\u3002<\/p>\n<ul>\n<li><strong>\u52a0\u8f7d\u73af\u5883<\/strong>:\n<pre><code>import verifiers as vf\r\n# \u52a0\u8f7d\u5df2\u5b89\u88c5\u7684\u73af\u5883\uff0c\u5e76\u4f20\u5165\u5fc5\u8981\u7684\u53c2\u6570\r\nvf_env = vf.load_environment(\"my-new-env\", **env_args)\r\n<\/code><\/pre>\n<\/li>\n<li><strong>\u5feb\u901f\u8bc4\u4f30\u73af\u5883<\/strong>: \u4f7f\u7528\u00a0<code>vf-eval<\/code>\u00a0\u547d\u4ee4\u53ef\u4ee5\u5feb\u901f\u6d4b\u8bd5\u4f60\u7684\u73af\u5883\u3002\u5b83\u9ed8\u8ba4\u4f7f\u7528\u00a0<code>gpt-4.1-mini<\/code>\u00a0\u6a21\u578b\uff0c\u5bf95\u4e2a\u63d0\u793a\u5404\u8fdb\u884c3\u6b21 rollout\u3002\n<pre><code># \u5bf9\u540d\u4e3a my-new-env \u7684\u73af\u5883\u8fdb\u884c\u8bc4\u4f30\r\nvf-eval my-new-env\r\n<\/code><\/pre>\n<\/li>\n<\/ul>\n<h3>4. \u73af\u5883\u7684\u6838\u5fc3\u5143\u7d20<\/h3>\n<p>\u4e00\u4e2a Verifiers \u73af\u5883\u4e3b\u8981\u7531\u4ee5\u4e0b\u51e0\u90e8\u5206\u6784\u6210\uff1a<\/p>\n<ul>\n<li><strong>\u6570\u636e\u96c6 (Datasets)<\/strong>: \u4e00\u4e2a Hugging Face \u6570\u636e\u96c6\uff0c\u5fc5\u987b\u5305\u542b\u4e00\u4e2a\u00a0<code>prompt<\/code>\u00a0\u5217\u4f5c\u4e3a\u8f93\u5165\u3002<\/li>\n<li><strong>\u4ea4\u4e92\u903b\u8f91 (Rollout logic)<\/strong>: \u6a21\u578b\u4e0e\u73af\u5883\u4e4b\u95f4\u7684\u4ea4\u4e92\u65b9\u5f0f\uff0c\u4f8b\u5982\u5728\u00a0<code>MultiTurnEnv<\/code>\u00a0\u4e2d\u5b9a\u4e49\u7684\u00a0<code>env_response<\/code>\u00a0\u548c\u00a0<code>is_completed<\/code>\u00a0\u65b9\u6cd5\u3002<\/li>\n<li><strong>\u8bc4\u4f30\u6807\u51c6 (Rubrics)<\/strong>: \u7528\u4e8e\u5c01\u88c5\u4e00\u4e2a\u6216\u591a\u4e2a\u5956\u52b1\u51fd\u6570\uff0c\u5bf9\u6a21\u578b\u7684\u8f93\u51fa\u8fdb\u884c\u8bc4\u5206\u3002<\/li>\n<li><strong>\u89e3\u6790\u5668 (Parsers)<\/strong>: \u53ef\u9009\u7ec4\u4ef6\uff0c\u7528\u4e8e\u5c01\u88c5\u53ef\u91cd\u7528\u7684\u89e3\u6790\u903b\u8f91\u3002<\/li>\n<\/ul>\n<h3>5. \u8bad\u7ec3\u6a21\u578b<\/h3>\n<p>Verifiers \u63d0\u4f9b\u4e86\u4e24\u79cd\u4e3b\u8981\u7684\u8bad\u7ec3\u65b9\u5f0f\uff1a<\/p>\n<ul>\n<li><strong>\u4f7f\u7528\u5185\u7f6e\u7684\u00a0<code>GRPOTrainer<\/code><\/strong>:<br \/>\n\u8fd9\u4e2a\u8bad\u7ec3\u5668\u9002\u7528\u4e8e\u5728 2-16 \u4e2a GPU \u4e0a\u9ad8\u6548\u8bad\u7ec3\u7a20\u5bc6 <a href=\"https:\/\/www.kdjingpai.com\/en\/transformer\/\">Transformer<\/a> \u6a21\u578b\u3002<\/p>\n<pre><code># \u6b65\u9aa41: \u542f\u52a8 vLLM \u63a8\u7406\u670d\u52a1\u5668 (shell 0)\r\n# \u5047\u8bbe\u4f7f\u75287\u4e2aGPU\u8fdb\u884c\u6570\u636e\u5e76\u884c\r\nCUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6 vf-vllm --model your-model-name \\\r\n--data-parallel-size 7 --enforce-eager --disable-log-requests\r\n# \u6b65\u9aa42: \u542f\u52a8\u8bad\u7ec3\u811a\u672c (shell 1)\r\n# \u4f7f\u7528\u5269\u4f59\u7684GPU\u8fdb\u884c\u8bad\u7ec3\r\nCUDA_VISIBLE_DEVICES=7 accelerate launch --num-processes 1 \\\r\n--config-file configs\/zero3.yaml examples\/grpo\/train_script.py --size 1.7B\r\n<\/code><\/pre>\n<\/li>\n<li><strong>\u4f7f\u7528\u00a0<code>prime-rl<\/code>\u00a0(\u63a8\u8350)<\/strong>:<br \/>\n<code>prime-rl<\/code>\u00a0\u662f\u4e00\u4e2a\u5916\u90e8\u9879\u76ee\uff0c\u5b83\u539f\u751f\u652f\u6301\u7528 Verifiers \u521b\u5efa\u7684\u73af\u5883\uff0c\u5e76\u901a\u8fc7 FSDP \u63d0\u4f9b\u4e86\u66f4\u597d\u7684\u6027\u80fd\u548c\u6269\u5c55\u6027\u3002 \u5b83\u7684\u914d\u7f6e\u548c\u7528\u6237\u4f53\u9a8c\u66f4\u6210\u719f\u3002<\/p>\n<pre><code># \u5728 prime-rl \u7684\u914d\u7f6e\u6587\u4ef6\u4e2d\u6307\u5b9a\u73af\u5883\r\n# orch.toml\r\n[environment]\r\nid = \"your-env-name\"\r\n# \u542f\u52a8 prime-rl \u8bad\u7ec3\r\nuv run rl \\\r\n--trainer @ configs\/your_exp\/train.toml \\\r\n--orchestrator @ configs\/your_exp\/orch.toml \\\r\n--inference @ configs\/your_exp\/infer.toml\r\n<\/code><\/pre>\n<\/li>\n<\/ul>\n<h2>\u5e94\u7528\u573a\u666f<\/h2>\n<ol>\n<li><strong>\u8bad\u7ec3\u7279\u5b9a\u4efb\u52a1\u7684\u667a\u80fd\u4f53<\/strong><br \/>\n\u5229\u7528\u00a0<code>ToolEnv<\/code>\u00a0\u6216\u00a0<code>MultiTurnEnv<\/code>\uff0c\u5f00\u53d1\u8005\u53ef\u4ee5\u521b\u5efa\u590d\u6742\u7684\u4ea4\u4e92\u73af\u5883\uff0c\u8bad\u7ec3 LLM \u667a\u80fd\u4f53\u5b66\u4e60\u5982\u4f55\u4f7f\u7528\u5916\u90e8\u5de5\u5177\uff08\u5982\u8ba1\u7b97\u5668\u3001\u641c\u7d22\u5f15\u64ce\uff09\u6216\u5728\u591a\u8f6e\u5bf9\u8bdd\u4e2d\u5b8c\u6210\u7279\u5b9a\u4efb\u52a1\uff08\u5982\u9884\u8ba2\u673a\u7968\u3001\u5ba2\u6237\u652f\u6301\uff09\u3002<\/li>\n<li><strong>\u6784\u5efa\u81ea\u52a8\u5316\u8bc4\u4f30\u6d41\u7a0b<\/strong><br \/>\n<code>SingleTurnEnv<\/code>\u00a0\u53ef\u4ee5\u7528\u6765\u6784\u5efa\u81ea\u52a8\u5316\u7684\u8bc4\u4f30\u6d41\u7a0b\u3002\u901a\u8fc7\u5b9a\u4e49\u4e00\u4e2a\u5305\u542b\u6807\u51c6\u7b54\u6848\u548c\u8bc4\u4f30\u6807\u51c6\uff08<code>Rubric<\/code>\uff09\u7684\u73af\u5883\uff0c\u53ef\u4ee5\u5bf9\u4e0d\u540c\u6a21\u578b\u7684\u6027\u80fd\u8fdb\u884c\u91cf\u5316\u6bd4\u8f83\uff0c\u4f8b\u5982\u8bc4\u4f30\u4ee3\u7801\u751f\u6210\u4efb\u52a1\u7684\u6b63\u786e\u7387\u6216\u6587\u672c\u6458\u8981\u7684\u8d28\u91cf\u3002<\/li>\n<li><strong>\u751f\u6210\u9ad8\u8d28\u91cf\u7684\u5408\u6210\u6570\u636e<\/strong><br \/>\n\u901a\u8fc7\u73af\u5883\u4ea4\u4e92\uff08rollout\uff09\u8fc7\u7a0b\uff0c\u53ef\u4ee5\u751f\u6210\u5927\u91cf\u6a21\u578b\u4e0e\u73af\u5883\u4ea4\u4e92\u7684\u6570\u636e\u3002\u8fd9\u4e9b\u6570\u636e\u53ef\u4ee5\u88ab\u4fdd\u5b58\u4e3a Hugging Face \u6570\u636e\u96c6\uff0c\u5e76\u7528\u4e8e\u540e\u7eed\u7684\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\u6216\u5176\u4ed6\u6a21\u578b\u7684\u8bad\u7ec3\uff0c\u8fd9\u662f\u4e00\u79cd\u9ad8\u6548\u7684\u5408\u6210\u6570\u636e\u751f\u6210\u7ba1\u9053\u3002<\/li>\n<li><strong>\u5b66\u672f\u7814\u7a76\u4e0e\u7b97\u6cd5\u9a8c\u8bc1<\/strong><br \/>\nVerifiers \u4e3a\u5f3a\u5316\u5b66\u4e60\u7814\u7a76\u8005\u63d0\u4f9b\u4e86\u4e00\u4e2a\u6a21\u5757\u5316\u3001\u53ef\u590d\u73b0\u7684\u5b9e\u9a8c\u5e73\u53f0\u3002\u7814\u7a76\u4eba\u5458\u53ef\u4ee5\u65b9\u4fbf\u5730\u5b9e\u73b0\u65b0\u7684\u4ea4\u4e92\u534f\u8bae\u3001\u5956\u52b1\u51fd\u6570\u6216\u8bad\u7ec3\u7b97\u6cd5\uff0c\u5e76\u5728\u6807\u51c6\u5316\u7684\u73af\u5883\u4e2d\u9a8c\u8bc1\u5176\u6709\u6548\u6027\u3002<\/li>\n<\/ol>\n<h2>QA<\/h2>\n<ol>\n<li><strong>Verifiers \u5e93\u548c prime-rl \u6709\u4ec0\u4e48\u5173\u7cfb\uff1f<\/strong><br \/>\n<code>prime-rl<\/code>\u00a0\u662f\u4e00\u4e2a\u72ec\u7acb\u7684\u8bad\u7ec3\u6846\u67b6\uff0c\u5b83\u539f\u751f\u652f\u6301\u4f7f\u7528 Verifiers \u521b\u5efa\u7684\u73af\u5883\u3002 Verifiers \u4e13\u6ce8\u4e8e\u63d0\u4f9b\u6784\u5efa RL \u73af\u5883\u7684\u7ec4\u4ef6\uff0c\u800c\u00a0<code>prime-rl<\/code>\u00a0\u5219\u4e13\u6ce8\u4e8e\u63d0\u4f9b\u4e00\u4e2a\u66f4\u5f3a\u5927\u3001\u6027\u80fd\u66f4\u4f18\u3001\u6269\u5c55\u6027\u66f4\u597d\u7684 FSDP\uff08\u5b8c\u5168\u5206\u7247\u6570\u636e\u5e76\u884c\uff09\u8bad\u7ec3\u65b9\u6848\u3002\u5bf9\u4e8e\u5927\u89c4\u6a21\u8bad\u7ec3\uff0c\u5b98\u65b9\u63a8\u8350\u4f7f\u7528\u00a0<code>prime-rl<\/code>\u3002<\/li>\n<li><strong>\u5982\u4f55\u4e3a\u6211\u7684\u73af\u5883\u5b9a\u4e49\u5956\u52b1\u51fd\u6570\uff1f<\/strong><br \/>\n\u4f60\u9700\u8981\u5728\u00a0<code>vf.Rubric<\/code>\u00a0\u5bf9\u8c61\u4e2d\u5b9a\u4e49\u4e00\u4e2a\u6216\u591a\u4e2a\u5956\u52b1\u51fd\u6570\u3002\u6bcf\u4e2a\u51fd\u6570\u63a5\u6536\u00a0<code>prompt<\/code>\u3001<code>completion<\/code>\u00a0\u7b49\u53c2\u6570\uff0c\u5e76\u8fd4\u56de\u4e00\u4e2a\u6d6e\u70b9\u6570\u4f5c\u4e3a\u5956\u52b1\u503c\u3002\u4f60\u8fd8\u53ef\u4ee5\u4e3a\u4e0d\u540c\u7684\u5956\u52b1\u51fd\u6570\u8bbe\u7f6e\u4e0d\u540c\u7684\u6743\u91cd\u3002<\/li>\n<li><strong>\u6211\u662f\u5426\u9700\u8981\u81ea\u5df1\u5b9e\u73b0\u6a21\u578b\u7684\u4ea4\u4e92\u903b\u8f91\uff1f<\/strong><br \/>\n\u4e0d\u4e00\u5b9a\u3002\u5bf9\u4e8e\u5355\u8f6e\u95ee\u7b54\u548c\u6807\u51c6\u5de5\u5177\u8c03\u7528\u573a\u666f\uff0c\u4f60\u53ef\u4ee5\u76f4\u63a5\u4f7f\u7528\u00a0<code>SingleTurnEnv<\/code>\u00a0\u548c\u00a0<code>ToolEnv<\/code>\u3002 \u53ea\u6709\u5f53\u4f60\u7684\u5e94\u7528\u9700\u8981\u975e\u5e38\u72ec\u7279\u7684\u3001\u975e\u6807\u51c6\u7684\u4ea4\u4e92\u6d41\u7a0b\u65f6\uff0c\u624d\u9700\u8981\u7ee7\u627f\u00a0<code>MultiTurnEnv<\/code>\u00a0\u5e76\u91cd\u5199\u00a0<code>is_completed<\/code>\u00a0\u548c\u00a0<code>env_response<\/code>\u00a0\u65b9\u6cd5\u3002<\/li>\n<li><strong>\u8bad\u7ec3\u65f6\u9047\u5230 NCCL \u76f8\u5173\u7684\u9519\u8bef\u600e\u4e48\u529e\uff1f<\/strong><br \/>\n\u6839\u636e\u5b98\u65b9\u6587\u6863\uff0cvLLM \u5728\u540c\u6b65\u6743\u91cd\u65f6\u53ef\u80fd\u4f1a\u9047\u5230 GPU \u95f4\u901a\u4fe1\u6302\u8d77\u7684\u95ee\u9898\u3002\u4f60\u53ef\u4ee5\u5c1d\u8bd5\u5728\u73af\u5883\u4e2d\u8bbe\u7f6e\u00a0<code>NCCL_P2P_DISABLE=1<\/code>\u00a0\u6765\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\u3002\u5982\u679c\u95ee\u9898\u4ecd\u7136\u5b58\u5728\uff0c\u53ef\u4ee5\u5c1d\u8bd5\u8bbe\u7f6e\u00a0<code>NCCL_CUMEM_ENABLE=1<\/code>\u00a0\u6216\u5411\u9879\u76ee\u63d0\u51fa\u4e00\u4e2a issue\u3002<\/li>\n<\/ol>\n","protected":false},"excerpt":{"rendered":"<p>Verifiers \u662f\u4e00\u4e2a\u7528\u4e8e\u521b\u5efa\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u73af\u5883\u548c\u8bad\u7ec3\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u4ee3\u7406\u7684\u6a21\u5757\u5316\u7ec4\u4ef6\u5e93\u3002 \u8fd9\u4e2a\u9879\u76ee\u7684\u76ee\u6807\u662f\u63d0\u4f9b\u4e00\u5957\u53ef\u9760\u7684\u5de5\u5177\uff0c\u8ba9\u5f00\u53d1\u8005\u53ef\u4ee5\u65b9\u4fbf\u5730\u6784\u5efa\u3001\u8bad\u7ec3\u548c\u8bc4\u4f30LLM\u4ee3\u7406\u3002 Verifiers \u5305\u542b\u4e00\u4e2a\u57fa\u4e8e\u00a0transforme&#8230;<\/p>\n","protected":false},"author":1,"featured_media":32782,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[425,20,483],"tags":[230,365],"class_list":["post-51114","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai-professional","category-tool","category-fine-tuning","tag-aikaiyuanxiangmu","tag-damoxingweidiao"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts\/51114","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/comments?post=51114"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts\/51114\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/media\/32782"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/media?parent=51114"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/categories?post=51114"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/tags?post=51114"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}