{"id":29685,"date":"2025-03-31T23:05:02","date_gmt":"2025-03-31T15:05:02","guid":{"rendered":"https:\/\/www.aisharenet.com\/?p=29685"},"modified":"2025-03-31T23:05:02","modified_gmt":"2025-03-31T15:05:02","slug":"dayuyanmoxingtuibei","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/de\/dayuyanmoxingtuibei\/","title":{"rendered":"\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\uff1a\u5728\u201c\u601d\u8003\u4e0d\u8db3\u201d\u4e0e\u201c\u8fc7\u5ea6\u601d\u8003\u201d\u4e4b\u95f4\u5bfb\u6c42\u5e73\u8861"},"content":{"rendered":"<p>\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u7684\u53d1\u5c55\u65e5\u65b0\u6708\u5f02\uff0c\u5176\u63a8\u7406\u80fd\u529b\u5df2\u6210\u4e3a\u8861\u91cf\u5176\u667a\u80fd\u6c34\u5e73\u7684\u5173\u952e\u6307\u6807\u3002\u7279\u522b\u662f\u5177\u5907\u957f\u63a8\u7406\u80fd\u529b\u7684\u6a21\u578b\uff0c\u4f8b\u5982 OpenAI \u7684\u00a0<code>o1<\/code>\u3001<code><a href=\"https:\/\/www.kdjingpai.com\/ja\/deepseek-r1nenglixiang\/\">DeepSeek-R1<\/a><\/code>\u3001<code><a href=\"https:\/\/www.kdjingpai.com\/ja\/xiaomoxingdanengba\/\">QwQ-32B<\/a><\/code>\u00a0\u548c\u00a0<code><a href=\"https:\/\/www.kdjingpai.com\/ja\/kimi-ai\/\">Kimi<\/a> K1.5<\/code>\u00a0\u7b49\uff0c\u5b83\u4eec\u901a\u8fc7\u6a21\u62df\u4eba\u7c7b\u6df1\u5ea6\u601d\u8003\u8fc7\u7a0b\u6765\u89e3\u51b3\u590d\u6742\u95ee\u9898\uff0c\u5f15\u53d1\u4e86\u5e7f\u6cdb\u5173\u6ce8\u3002\u8fd9\u79cd\u80fd\u529b\u901a\u5e38\u6d89\u53ca\u5230\u4e00\u79cd\u79f0\u4e3a\u201c\u63a8\u7406\u65f6\u6269\u5c55\u201d\uff08Inference-Time Scaling\uff09\u7684\u6280\u672f\uff0c\u5141\u8bb8\u6a21\u578b\u5728\u751f\u6210\u7b54\u6848\u7684\u8fc7\u7a0b\u4e2d\uff0c\u6295\u5165\u66f4\u591a\u65f6\u95f4\u8fdb\u884c\u63a2\u7d22\u548c\u4fee\u6b63\u3002<\/p>\n<p>\u7136\u800c\uff0c\u6df1\u5165\u7814\u7a76\u53d1\u73b0\uff0c\u8fd9\u4e9b\u6a21\u578b\u5728\u63a8\u7406\u65f6\u5e38\u9677\u5165\u4e24\u4e2a\u6781\u7aef\uff1a<strong>\u601d\u8003\u4e0d\u8db3 (Underthinking)<\/strong>\u00a0\u548c\u00a0<strong>\u8fc7\u5ea6\u601d\u8003 (Overthinking)<\/strong>\u3002<\/p>\n<p><strong>\u601d\u8003\u4e0d\u8db3<\/strong>\u00a0\u6307\u7684\u662f\u6a21\u578b\u5728\u63a8\u7406\u4e2d\u9891\u7e41\u5207\u6362\u601d\u8def\uff0c\u96be\u4ee5\u96c6\u4e2d\u4e8e\u4e00\u4e2a\u6709\u524d\u666f\u7684\u65b9\u5411\u8fdb\u884c\u6df1\u5165\u6316\u6398\u3002\u6a21\u578b\u8f93\u51fa\u4e2d\u53ef\u80fd\u5145\u65a5\u7740 \u201calternatively\u201d\u3001\u201cbut wait\u201d\u3001\u201clet me reconsider\u201d \u7b49\u8bcd\u8bed\uff0c\u5982\u540c\u4e0b\u56fe\u6240\u793a\uff0c\u5bfc\u81f4\u6700\u7ec8\u7b54\u6848\u9519\u8bef\u3002\u8fd9\u79cd\u73b0\u8c61\u53ef\u7c7b\u6bd4\u4e3a\u4eba\u7c7b\u6ce8\u610f\u529b\u4e0d\u96c6\u4e2d\uff0c\u5f71\u54cd\u4e86\u63a8\u7406\u7684\u6709\u6548\u6027\u3002<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/a4c24f95239954b.png\" alt=\"\u6a21\u578b\u601d\u8003\u4e0d\u8db3\u793a\u4f8b\" width=\"664\" height=\"364\" \/><\/p>\n<p><strong>\u8fc7\u5ea6\u601d\u8003<\/strong>\u00a0\u5219\u8868\u73b0\u4e3a\u6a21\u578b\u5728\u7b80\u5355\u95ee\u9898\u4e0a\u751f\u6210\u5197\u957f\u800c\u4e0d\u5fc5\u8981\u7684\u201c\u601d\u7ef4\u94fe\u201d\u3002\u4f8b\u5982\uff0c\u5bf9\u4e8e \u201c2+3=\uff1f\u201d \u8fd9\u6837\u57fa\u7840\u7684\u7b97\u672f\u9898\uff0c\u67d0\u4e9b\u6a21\u578b\u53ef\u80fd\u8017\u8d39\u6570\u767e\u751a\u81f3\u4e0a\u5343\u00a0<code><a href=\"https:\/\/www.kdjingpai.com\/ja\/tokenization\/\">token<\/a><\/code>\u00a0\u6765\u53cd\u590d\u9a8c\u8bc1\u6216\u63a2\u7d22\u591a\u79cd\u89e3\u6cd5\uff0c\u5982\u4e0b\u56fe\u6240\u793a\u3002\u867d\u7136\u590d\u6742\u7684\u601d\u7ef4\u8fc7\u7a0b\u5bf9\u96be\u9898\u6709\u76ca\uff0c\u4f46\u5728\u7b80\u5355\u573a\u666f\u4e0b\uff0c\u8fd9\u65e0\u7591\u9020\u6210\u4e86\u8ba1\u7b97\u8d44\u6e90\u7684\u6d6a\u8d39\u3002<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/932ea4eee59c1b7.png\" alt=\"\u6a21\u578b\u8fc7\u5ea6\u601d\u8003\u793a\u4f8b\" width=\"672\" height=\"626\" \/><\/p>\n<p>\u8fd9\u4e24\u4e2a\u95ee\u9898\u5171\u540c\u6307\u5411\u4e86\u4e00\u4e2a\u6838\u5fc3\u6311\u6218\uff1a\u5982\u4f55\u5728\u4fdd\u8bc1\u7b54\u6848\u8d28\u91cf\u7684\u524d\u63d0\u4e0b\uff0c\u63d0\u5347\u6a21\u578b\u7684\u601d\u8003\u6548\u7387\uff1f\u7406\u60f3\u7684\u6a21\u578b\u5e94\u5f53\u80fd\u5728\u6700\u77ed\u7684\u8f93\u51fa\u5185\uff0c\u627e\u5230\u5e76\u7ed9\u51fa\u6b63\u786e\u7b54\u6848\u3002<\/p>\n<p>\u4e3a\u4e86\u5e94\u5bf9\u8fd9\u4e00\u6311\u6218\uff0c<code>EvalScope<\/code>\u00a0\u9879\u76ee\u5f15\u5165\u4e86\u00a0<code>EvalThink<\/code>\u00a0\u7ec4\u4ef6\uff0c\u65e8\u5728\u63d0\u4f9b\u4e00\u4e2a\u6807\u51c6\u5316\u7684\u5de5\u5177\u6765\u8bc4\u4f30\u6a21\u578b\u7684\u601d\u8003\u6548\u7387\u3002\u672c\u6587\u5c06\u4ee5\u00a0<code>MATH-500<\/code>\u00a0\u6570\u636e\u96c6\u4e3a\u4f8b\uff0c\u5206\u6790\u5305\u62ec\u00a0<code>DeepSeek-R1-Distill-Qwen-7B<\/code>\u00a0\u5728\u5185\u7684\u4e00\u7cfb\u5217\u63a8\u7406\u6a21\u578b\u7684\u8868\u73b0\uff0c\u91cd\u70b9\u8003\u5bdf\u516d\u4e2a\u7ef4\u5ea6\uff1a\u6a21\u578b\u63a8\u7406\u00a0<code>token<\/code>\u00a0\u6570\u3001\u9996\u6b21\u6b63\u786e\u00a0<code>token<\/code>\u00a0\u6570\u3001\u5269\u4f59\u53cd\u601d\u00a0<code>token<\/code>\u00a0\u6570\u3001<code>token<\/code>\u00a0\u6548\u7387\u3001\u5b50\u601d\u7ef4\u94fe\u6570\u91cf\u548c\u51c6\u786e\u7387\u3002<\/p>\n<p>&nbsp;<\/p>\n<h2>\u8bc4\u4f30\u65b9\u6cd5\u4e0e\u6d41\u7a0b<\/h2>\n<p>\u8bc4\u4f30\u8fc7\u7a0b\u4e3b\u8981\u5305\u542b\u4e24\u4e2a\u9636\u6bb5\uff1a\u6a21\u578b\u63a8\u7406\u8bc4\u4f30\u548c\u6a21\u578b\u601d\u8003\u6548\u7387\u8bc4\u6d4b\u3002<\/p>\n<h3>\u6a21\u578b\u63a8\u7406\u8bc4\u4f30<\/h3>\n<p>\u6b64\u9636\u6bb5\u7684\u76ee\u6807\u662f\u83b7\u53d6\u6a21\u578b\u5728\u00a0<code>MATH-500<\/code>\u00a0\u6570\u636e\u96c6\u4e0a\u7684\u539f\u59cb\u63a8\u7406\u7ed3\u679c\u548c\u57fa\u7840\u51c6\u786e\u7387\u3002<code>MATH-500<\/code>\u00a0\u6570\u636e\u96c6\u5305\u542b 500 \u4e2a\u4e0d\u540c\u96be\u5ea6\u7684\u6570\u5b66\u95ee\u9898\uff08\u4ece Level 1 \u5230 Level 5\uff09\u3002<\/p>\n<p><strong>\u51c6\u5907\u8bc4\u6d4b\u73af\u5883<\/strong><\/p>\n<p>\u8bc4\u4f30\u53ef\u901a\u8fc7\u63a5\u5165\u517c\u5bb9 OpenAI API \u7684\u63a8\u7406\u670d\u52a1\u8fdb\u884c\u3002<code>EvalScope<\/code>\u00a0\u6846\u67b6\u4e5f\u652f\u6301\u4f7f\u7528\u00a0<code>transformers<\/code>\u00a0\u5e93\u5728\u672c\u5730\u8fdb\u884c\u8bc4\u6d4b\u3002\u5bf9\u4e8e\u9700\u8981\u5904\u7406\u957f\u601d\u7ef4\u94fe\uff08\u53ef\u80fd\u8d85\u8fc7 10,000\u00a0<code>token<\/code>\uff09\u7684\u63a8\u7406\u6a21\u578b\uff0c\u4f7f\u7528\u00a0<code><a href=\"https:\/\/www.kdjingpai.com\/ja\/vllm\/\">vLLM<\/a><\/code>\u00a0\u6216\u00a0<code><a href=\"https:\/\/www.kdjingpai.com\/ja\/ollama\/\">ollama<\/a><\/code>\u00a0\u7b49\u9ad8\u6548\u63a8\u7406\u6846\u67b6\u90e8\u7f72\u6a21\u578b\uff0c\u53ef\u4ee5\u663e\u8457\u52a0\u901f\u8bc4\u6d4b\u8fc7\u7a0b\u3002<\/p>\n<p>\u4ee5\u00a0<code>DeepSeek-R1-Distill-Qwen-7B<\/code>\u00a0\u4e3a\u4f8b\uff0c\u4f7f\u7528\u00a0<code>vLLM<\/code>\u00a0\u90e8\u7f72\u670d\u52a1\u7684\u793a\u4f8b\u547d\u4ee4\u5982\u4e0b\uff1a<\/p>\n<pre><code>VLLM_USE_MODELSCOPE=True CUDA_VISIBLE_DEVICES=0 python -m vllm.entrypoints.openai.api_server --model deepseek-ai\/DeepSeek-R1-Distill-Qwen-7B --served-model-name DeepSeek-R1-Distill-Qwen-7B --trust_remote_code --port 8801\r\n<\/code><\/pre>\n<p><strong>\u6267\u884c\u63a8\u7406\u8bc4\u6d4b<\/strong><\/p>\n<p>\u901a\u8fc7\u00a0<code>EvalScope<\/code>\u00a0\u7684\u00a0<code>TaskConfig<\/code>\u00a0\u914d\u7f6e\u6a21\u578b API \u5730\u5740\u3001\u540d\u79f0\u3001\u6570\u636e\u96c6\u3001\u6279\u5904\u7406\u5927\u5c0f\u548c\u751f\u6210\u53c2\u6570\uff0c\u7136\u540e\u8fd0\u884c\u8bc4\u6d4b\u4efb\u52a1\u3002\u4ee5\u4e0b\u4e3a\u793a\u4f8b Python \u4ee3\u7801\uff1a<\/p>\n<pre><code>from evalscope import TaskConfig, run_task\r\ntask_config = TaskConfig(\r\napi_url='http:\/\/0.0.0.0:8801\/v1\/chat\/completions',  # \u63a8\u7406\u670d\u52a1\u5730\u5740\r\nmodel='DeepSeek-R1-Distill-Qwen-7B',  # \u6a21\u578b\u540d\u79f0 (\u9700\u4e0e\u90e8\u7f72\u65f6\u4e00\u81f4)\r\neval_type='service',  # \u8bc4\u6d4b\u7c7b\u578b\uff1a\u670d\u52a1\r\ndatasets=['math_500'],  # \u6570\u636e\u96c6\r\ndataset_args={'math_500': {'few_shot_num': 0, 'subset_list': ['Level 1', 'Level 2', 'Level 3', 'Level 4', 'Level 5']}},  # \u6570\u636e\u96c6\u53c2\u6570\uff0c\u5305\u542b\u96be\u5ea6\u7ea7\u522b\r\neval_batch_size=32,  # \u5e76\u53d1\u8bf7\u6c42\u6570\r\ngeneration_config={\r\n'max_tokens': 20000,  # \u6700\u5927\u751f\u6210 token \u6570\uff0c\u8bbe\u7f6e\u8f83\u5927\u503c\u9632\u622a\u65ad\r\n'temperature': 0.6,  # \u91c7\u6837\u6e29\u5ea6\r\n'top_p': 0.95,  # top-p \u91c7\u6837\r\n'n': 1,  # \u6bcf\u4e2a\u8bf7\u6c42\u751f\u6210\u4e00\u4e2a\u56de\u590d\r\n},\r\n)\r\nrun_task(task_config)\r\n<\/code><\/pre>\n<p>\u8bc4\u6d4b\u5b8c\u6210\u540e\uff0c\u4f1a\u8f93\u51fa\u6a21\u578b\u5728\u00a0<code>MATH-500<\/code>\u00a0\u5404\u96be\u5ea6\u7ea7\u522b\u4e0a\u7684\u51c6\u786e\u7387\uff08<code>AveragePass@1<\/code>\uff09\uff1a<\/p>\n<pre><code>| Model                       | Dataset   | Metric        | Subset   | Num | Score  | Cat.0   |\r\n|-----------------------------|-----------|---------------|----------|-----|--------|---------|\r\n| DeepSeek-R1-Distill-Qwen-7B | math_500  | AveragePass@1 | Level 1  | 43  | 0.9535 | default |\r\n| DeepSeek-R1-Distill-Qwen-7B | math_500  | AveragePass@1 | Level 2  | 90  | 0.9667 | default |\r\n| DeepSeek-R1-Distill-Qwen-7B | math_500  | AveragePass@1 | Level 3  | 105 | 0.9587 | default |\r\n| DeepSeek-R1-Distill-Qwen-7B | math_500  | AveragePass@1 | Level 4  | 128 | 0.9115 | default |\r\n| DeepSeek-R1-Distill-Qwen-7B | math_500  | AveragePass@1 | Level 5  | 134 | 0.8557 | default |\r\n<\/code><\/pre>\n<h3>\u6a21\u578b\u601d\u8003\u6548\u7387\u8bc4\u4f30<\/h3>\n<p>\u83b7\u53d6\u63a8\u7406\u7ed3\u679c\u540e\uff0c<code>EvalThink<\/code>\u00a0\u7ec4\u4ef6\u4ecb\u5165\uff0c\u8fdb\u884c\u66f4\u6df1\u5165\u7684\u6548\u7387\u5206\u6790\u3002\u6838\u5fc3\u8bc4\u4f30\u6307\u6807\u5305\u62ec\uff1a<\/p>\n<ul>\n<li><strong>\u6a21\u578b\u63a8\u7406\u00a0<code>token<\/code>\u00a0\u6570 (Reasoning Tokens)<\/strong>: \u6a21\u578b\u751f\u6210\u7b54\u6848\u8fc7\u7a0b\u4e2d\uff0c\u601d\u8003\u94fe\uff08\u5982 O1\/R1 \u6a21\u578b\u4e2d\u00a0<code>&lt;\/think&gt;<\/code>\u00a0\u6807\u5fd7\u524d\u7684\u5185\u5bb9\uff09\u6240\u5305\u542b\u7684\u00a0<code>token<\/code>\u00a0\u603b\u91cf\u3002<\/li>\n<li><strong>\u9996\u6b21\u6b63\u786e\u00a0<code>token<\/code>\u00a0\u6570 (First Correct Tokens)<\/strong>: \u4ece\u6a21\u578b\u8f93\u51fa\u5f00\u59cb\uff0c\u5230\u9996\u6b21\u51fa\u73b0\u53ef\u8bc6\u522b\u6b63\u786e\u7b54\u6848\u4f4d\u7f6e\u7684\u00a0<code>token<\/code>\u00a0\u6570\u91cf\u3002<\/li>\n<li><strong>\u5269\u4f59\u53cd\u601d\u00a0<code>token<\/code>\u00a0\u6570 (Reflection Tokens)<\/strong>: \u4ece\u9996\u6b21\u6b63\u786e\u7b54\u6848\u4f4d\u7f6e\u5230\u601d\u8003\u94fe\u7ed3\u675f\u7684\u00a0<code>token<\/code>\u00a0\u6570\u91cf\u3002\u8fd9\u90e8\u5206\u53cd\u6620\u4e86\u6a21\u578b\u627e\u5230\u7b54\u6848\u540e\u7ee7\u7eed\u9a8c\u8bc1\u6216\u63a2\u7d22\u6240\u82b1\u8d39\u7684\u4ee3\u4ef7\u3002<\/li>\n<li><strong>\u5b50\u601d\u7ef4\u94fe\u6570\u91cf (Num Thought)<\/strong>: \u901a\u8fc7\u7edf\u8ba1\u7279\u5b9a\u6807\u5fd7\u8bcd\uff08\u5982\u00a0<code>alternatively<\/code>,\u00a0<code>but wait<\/code>,\u00a0<code>let me reconsider<\/code>\uff09\u7684\u51fa\u73b0\u6b21\u6570\u6765\u4f30\u7b97\u6a21\u578b\u5207\u6362\u601d\u8def\u7684\u9891\u7387\u3002<\/li>\n<li><strong><code>token<\/code>\u00a0\u6548\u7387 (Token Efficiency)<\/strong>: \u8861\u91cf\u6709\u6548\u601d\u8003\u00a0<code>token<\/code>\u00a0\u5360\u6bd4\u7684\u6307\u6807\uff0c\u8ba1\u7b97\u516c\u5f0f\u4e3a\u9996\u6b21\u6b63\u786e\u00a0<code>token<\/code>\u00a0\u6570\u4e0e\u603b\u63a8\u7406\u00a0<code>token<\/code>\u00a0\u6570\u7684\u6bd4\u503c\u7684\u5e73\u5747\u503c\uff08\u4ec5\u8ba1\u5165\u56de\u7b54\u6b63\u786e\u7684\u6837\u672c\uff09\uff1a<br \/>\nToken Efficiency =\u00a0<sup>1<\/sup>\u2044<sub>N<\/sub>\u00a0\u2211\u00a0<sup>First Correct <a href=\"https:\/\/www.kdjingpai.com\/ja\/tokenization\/\">Tokens<\/a><sub>i<\/sub><\/sup>\u2044<sub>Reasoning Tokensi<\/sub><br \/>\n\u5176\u4e2d N \u4e3a\u56de\u7b54\u6b63\u786e\u7684\u95ee\u9898\u6570\u91cf\u3002\u8be5\u503c\u8d8a\u9ad8\uff0c\u8868\u793a\u6a21\u578b\u7684\u601d\u8003\u8d8a\u201c\u9ad8\u6548\u201d\u3002<\/li>\n<\/ul>\n<p>\u4e3a\u4e86\u786e\u5b9a\u201c\u9996\u6b21\u6b63\u786e\u00a0<code>token<\/code>\u00a0\u6570\u201d\uff0c\u8be5\u8bc4\u4f30\u6846\u67b6\u501f\u9274\u4e86\u00a0<code>ProcessBench<\/code>\u00a0\u7684\u601d\u8def\uff0c\u91c7\u7528\u4e00\u4e2a\u72ec\u7acb\u7684\u201c\u88c1\u5224\u201d\u6a21\u578b\uff08Judge Model\uff09\uff0c\u4f8b\u5982\u00a0<code>Qwen2.5-72B-Instruct<\/code>\uff0c\u6765\u68c0\u67e5\u63a8\u7406\u6b65\u9aa4\uff0c\u5b9a\u4f4d\u6700\u65e9\u51fa\u73b0\u6b63\u786e\u7b54\u6848\u7684\u4f4d\u7f6e\u3002\u5b9e\u73b0\u65b9\u5f0f\u6d89\u53ca\u5c06\u6a21\u578b\u8f93\u51fa\u6309\u6b65\u9aa4\u5206\u89e3\uff08\u7b56\u7565\u53ef\u9009\uff1a\u6309\u7279\u5b9a\u5206\u9694\u7b26\u00a0<code>separator<\/code>\u3001\u6309\u5173\u952e\u8bcd\u00a0<code>keywords<\/code>\u3001\u6216\u7531 LLM \u8f85\u52a9\u91cd\u5199\u5e76\u5207\u5206\u00a0<code>llm<\/code>\uff09\uff0c\u7136\u540e\u8ba9\u88c1\u5224\u6a21\u578b\u9010\u4e00\u5224\u65ad\u3002<\/p>\n<p>\u6267\u884c\u601d\u8003\u6548\u7387\u8bc4\u4f30\u7684\u793a\u4f8b\u4ee3\u7801\uff1a<\/p>\n<pre><code>from evalscope.third_party.thinkbench import run_task\r\n# \u914d\u7f6e\u88c1\u5224\u6a21\u578b\u670d\u52a1\r\njudge_config = dict(\r\napi_key='EMPTY',\r\nbase_url='http:\/\/0.0.0.0:8801\/v1', # \u5047\u8bbe\u88c1\u5224\u6a21\u578b\u4e5f\u90e8\u7f72\u5728\u6b64\u670d\u52a1\r\nmodel_name='Qwen2.5-72B-Instruct',\r\n)\r\n# \u914d\u7f6e\u5f85\u8bc4\u4f30\u6a21\u578b\u7684\u4fe1\u606f\r\nmodel_config = dict(\r\nreport_path='.\/outputs\/2025xxxx',  # \u4e0a\u4e00\u6b65\u63a8\u7406\u7ed3\u679c\u8def\u5f84\r\nmodel_name='DeepSeek-R1-Distill-Qwen-7B',  # \u6a21\u578b\u540d\u79f0\r\ntokenizer_path='deepseek-ai\/DeepSeek-R1-Distill-Qwen-7B',  # Tokenizer \u8def\u5f84\uff0c\u7528\u4e8e\u8ba1\u7b97 token\r\ndataset_name='math_500',  # \u6570\u636e\u96c6\u540d\u79f0\r\nsubsets=['Level 1', 'Level 2', 'Level 3', 'Level 4', 'Level 5'],  # \u6570\u636e\u96c6\u5b50\u96c6\r\nsplit_strategies='separator',  # \u63a8\u7406\u6b65\u9aa4\u5206\u5272\u7b56\u7565\r\njudge_config=judge_config\r\n)\r\nmax_tokens = 20000  # \u8fc7\u6ee4 token \u8fc7\u957f\u7684\u8f93\u51fa\r\ncount = 200  # \u6bcf\u4e2a\u5b50\u96c6\u62bd\u6837\u6570\u91cf\uff0c\u52a0\u901f\u8bc4\u6d4b\r\n# \u8fd0\u884c\u601d\u8003\u6548\u7387\u8bc4\u4f30\r\nrun_task(model_config, output_dir='outputs', max_tokens=max_tokens, count=count)\r\n<\/code><\/pre>\n<p>\u8bc4\u4f30\u7ed3\u679c\u4f1a\u8be6\u7ec6\u5217\u51fa\u6a21\u578b\u5728\u5404\u96be\u5ea6\u7ea7\u522b\u4e0a\u7684\u516d\u4e2a\u7ef4\u5ea6\u6307\u6807\u3002<\/p>\n<p>&nbsp;<\/p>\n<h2>\u7ed3\u679c\u5206\u6790\u4e0e\u8ba8\u8bba<\/h2>\n<p>\u7814\u7a76\u56e2\u961f\u4f7f\u7528\u00a0<code>EvalThink<\/code>\u00a0\u5bf9\u00a0<code>DeepSeek-R1-Distill-Qwen-7B<\/code>\u00a0\u53ca\u5176\u4ed6\u51e0\u4e2a\u6a21\u578b\uff08<code>QwQ-32B<\/code>\u3001<code>QwQ-32B-Preview<\/code>\u3001<code>DeepSeek-R1<\/code>\u3001<code>DeepSeek-R1-Distill-Qwen-32B<\/code>\uff09\u8fdb\u884c\u4e86\u8bc4\u4f30\uff0c\u5e76\u52a0\u5165\u4e86\u4e00\u4e2a\u975e\u63a8\u7406\u7684\u6570\u5b66\u4e13\u7528\u6a21\u578b\u00a0<code>Qwen2.5-Math-7B-Instruct<\/code>\u00a0\u4f5c\u4e3a\u5bf9\u6bd4\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/9bd6280f6b869d0.png\" alt=\"DeepSeek-R1-Distill-Qwen-7B \u601d\u8003\u6548\u7387\" \/><br \/>\n<em>\u56fe1\uff1aDeepSeek-R1-Distill-Qwen-7B \u601d\u8003\u6548\u7387\u6307\u6807<\/em><\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/e883bdcb1da5960.png\" alt=\"6\u4e2a\u6a21\u578b\u601d\u8003\u6548\u7387\u5bf9\u6bd4\" \/><br \/>\n<em>\u56fe2\uff1a6\u4e2a\u6a21\u578b\u5728 MATH-500 \u4e0d\u540c\u96be\u5ea6\u7ea7\u522b\u4e0a\u7684\u601d\u8003\u6548\u7387\u5bf9\u6bd4<\/em><\/p>\n<p>\u4ece\u5bf9\u6bd4\u7ed3\u679c\uff08\u56fe2\uff09\u4e2d\u53ef\u4ee5\u89c2\u5bdf\u5230\u4ee5\u4e0b\u8d8b\u52bf\uff1a<\/p>\n<ol>\n<li><strong>\u96be\u5ea6\u4e0e\u8868\u73b0\u5173\u8054<\/strong>: \u968f\u7740\u95ee\u9898\u96be\u5ea6\uff08Level 1 \u5230 Level 5\uff09\u589e\u52a0\uff0c\u5927\u90e8\u5206\u6a21\u578b\u51c6\u786e\u7387\u4e0b\u964d\u3002\u7136\u800c\uff0c<code>QwQ-32B<\/code>\u00a0\u548c\u00a0<code>DeepSeek-R1<\/code>\u00a0\u5728\u9ad8\u96be\u5ea6\u95ee\u9898\u4e0a\u8868\u73b0\u7a81\u51fa\uff0c<code>QwQ-32B<\/code>\u00a0\u5728 Level 5 \u51c6\u786e\u7387\u6700\u9ad8\u3002\u540c\u65f6\uff0c\u6240\u6709\u6a21\u578b\u7684\u8f93\u51fa\u00a0<code>token<\/code>\u00a0\u6570\u90fd\u968f\u96be\u5ea6\u589e\u52a0\u800c\u53d8\u957f\uff0c\u8fd9\u7b26\u5408\u201c\u63a8\u7406\u65f6\u6269\u5c55\u201d\u7684\u9884\u671f\u2014\u2014\u6a21\u578b\u9700\u8981\u66f4\u591a\u201c\u601d\u8003\u201d\u6765\u89e3\u51b3\u96be\u9898\u3002<\/li>\n<li><strong>O1\/R1 \u7c7b\u63a8\u7406\u6a21\u578b\u7279\u6027<\/strong>:\n<ul>\n<li><strong>\u6548\u7387\u63d0\u5347<\/strong>: \u6709\u8da3\u7684\u662f\uff0c\u5bf9\u4e8e\u00a0<code>DeepSeek-R1<\/code>\u00a0\u548c\u00a0<code>QwQ-32B<\/code>\u00a0\u8fd9\u7c7b\u63a8\u7406\u6a21\u578b\uff0c\u867d\u7136\u8f93\u51fa\u53d8\u957f\uff0c\u4f46\u00a0<code>token<\/code>\u00a0\u6548\u7387\uff08\u6709\u6548\u00a0<code>token<\/code>\u00a0\u5360\u6bd4\uff09\u4e5f\u968f\u96be\u5ea6\u63d0\u5347\uff08<code>DeepSeek-R1<\/code>\u00a0\u4ece 36% \u5230 54%\uff0c<code>QwQ-32B<\/code>\u00a0\u4ece 31% \u5230 49%\uff09\u3002\u8fd9\u8868\u660e\u5b83\u4eec\u5728\u96be\u9898\u4e0a\u7684\u989d\u5916\u601d\u8003\u66f4\u5177\u201c\u6027\u4ef7\u6bd4\u201d\uff0c\u800c\u5728\u7b80\u5355\u95ee\u9898\u4e0a\u53ef\u80fd\u5b58\u5728\u4e00\u5b9a\u7684\u201c\u8fc7\u5ea6\u601d\u8003\u201d\uff0c\u4f8b\u5982\u4e0d\u5fc5\u8981\u7684\u53cd\u590d\u9a8c\u8bc1\u3002<code>QwQ-32B<\/code>\u00a0\u7684\u00a0<code>token<\/code>\u00a0\u6d88\u8017\u91cf\u6574\u4f53\u504f\u9ad8\uff0c\u8fd9\u6216\u8bb8\u662f\u5176\u80fd\u5728 Level 5 \u4fdd\u6301\u9ad8\u51c6\u786e\u7387\u7684\u539f\u56e0\u4e4b\u4e00\uff0c\u4f46\u4e5f\u6697\u793a\u4e86\u5176\u8fc7\u5ea6\u601d\u8003\u7684\u503e\u5411\u3002<\/li>\n<li><strong>\u601d\u7ef4\u8def\u5f84<\/strong>:\u00a0<code><a href=\"https:\/\/www.kdjingpai.com\/ja\/deepseek-chatshena\/\">DeepSeek<\/a><\/code>\u00a0\u7cfb\u5217\u6a21\u578b\u5728 Level 1-4 \u7684\u5b50\u601d\u7ef4\u94fe\u6570\u91cf\u76f8\u5bf9\u7a33\u5b9a\uff0c\u4f46\u5728\u6700\u96be\u7684 Level 5 \u6025\u5267\u589e\u52a0\uff0c\u8868\u660e Level 5 \u5bf9\u8fd9\u4e9b\u6a21\u578b\u6784\u6210\u4e86\u663e\u8457\u6311\u6218\uff0c\u9700\u8981\u591a\u6b21\u5c1d\u8bd5\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c<code>QwQ-32B<\/code>\u00a0\u7cfb\u5217\u6a21\u578b\u7684\u601d\u7ef4\u94fe\u6570\u91cf\u589e\u957f\u66f4\u5e73\u6ed1\uff0c\u53cd\u6620\u4e86\u4e0d\u540c\u7684\u5e94\u5bf9\u7b56\u7565\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u975e\u63a8\u7406\u6a21\u578b\u5c40\u9650<\/strong>: \u6570\u5b66\u4e13\u7528\u6a21\u578b\u00a0<code>Qwen2.5-Math-7B-Instruct<\/code>\u00a0\u5728\u5904\u7406\u9ad8\u96be\u5ea6\u95ee\u9898\u65f6\u51c6\u786e\u7387\u5927\u5e45\u4e0b\u964d\uff0c\u4e14\u5176\u8f93\u51fa\u00a0<code>token<\/code>\u00a0\u6570\u8fdc\u5c11\u4e8e\u63a8\u7406\u6a21\u578b\uff08\u7ea6\u4e09\u5206\u4e4b\u4e00\uff09\u3002\u8fd9\u663e\u793a\uff0c\u867d\u7136\u8fd9\u7c7b\u6a21\u578b\u5728\u666e\u901a\u95ee\u9898\u4e0a\u53ef\u80fd\u66f4\u5feb\u3001\u66f4\u7701\u8d44\u6e90\uff0c\u4f46\u7f3a\u4e4f\u6df1\u5ea6\u601d\u8003\u8fc7\u7a0b\u4f7f\u5176\u5728\u590d\u6742\u63a8\u7406\u4efb\u52a1\u4e0a\u5b58\u5728\u660e\u663e\u7684\u6027\u80fd\u201c\u5929\u82b1\u677f\u201d\u3002<\/li>\n<\/ol>\n<p>&nbsp;<\/p>\n<h2>\u65b9\u6cd5\u5b66\u8003\u91cf\u4e0e\u5c40\u9650\u6027<\/h2>\n<p>\u5728\u5e94\u7528\u00a0<code>EvalThink<\/code>\u00a0\u8fdb\u884c\u8bc4\u4f30\u65f6\uff0c\u9700\u8981\u6ce8\u610f\u4ee5\u4e0b\u51e0\u70b9\uff1a<\/p>\n<ul>\n<li><strong>\u6307\u6807\u5b9a\u4e49<\/strong>:\n<ul>\n<li>\u672c\u6587\u63d0\u51fa\u7684\u00a0<code>token<\/code>\u00a0\u6548\u7387\u6307\u6807\uff0c\u501f\u9274\u4e86\u6587\u732e\u4e2d\u5173\u4e8e\u201c\u8fc7\u5ea6\u601d\u8003\u201d\u548c\u201c\u601d\u8003\u4e0d\u8db3\u201d\u7684\u6982\u5ff5\uff0c\u4e3b\u8981\u5173\u6ce8\u00a0<code>token<\/code>\u00a0\u6570\u91cf\uff0c\u662f\u5bf9\u601d\u8003\u8fc7\u7a0b\u7684\u4e00\u79cd\u7b80\u5316\u5ea6\u91cf\uff0c\u672a\u80fd\u6355\u6349\u601d\u8003\u8d28\u91cf\u7684\u5168\u90e8\u7ec6\u8282\u3002<\/li>\n<li>\u5b50\u601d\u7ef4\u94fe\u6570\u91cf\u7684\u8ba1\u7b97\u4f9d\u8d56\u4e8e\u9884\u5b9a\u4e49\u7684\u5173\u952e\u8bcd\uff0c\u53ef\u80fd\u9700\u8981\u9488\u5bf9\u4e0d\u540c\u6a21\u578b\u8c03\u6574\u5173\u952e\u8bcd\u5217\u8868\u624d\u80fd\u51c6\u786e\u53cd\u6620\u5176\u601d\u8003\u6a21\u5f0f\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u9002\u7528\u8303\u56f4<\/strong>:\n<ul>\n<li>\u5f53\u524d\u6307\u6807\u4e3b\u8981\u5728\u6570\u5b66\u63a8\u7406\u6570\u636e\u96c6\u4e0a\u9a8c\u8bc1\uff0c\u5176\u5728\u5f00\u653e\u95ee\u7b54\u3001\u521b\u610f\u751f\u6210\u7b49\u5176\u4ed6\u573a\u666f\u7684\u6709\u6548\u6027\u6709\u5f85\u68c0\u9a8c\u3002<\/li>\n<li>\u8003\u8651\u5230\u00a0<code>DeepSeek-R1-Distill-Qwen-7B<\/code>\u00a0\u662f\u57fa\u4e8e\u6570\u5b66\u6a21\u578b\u84b8\u998f\u800c\u6765\uff0c\u5176\u5728\u00a0<code>MATH-500<\/code>\u00a0\u6570\u636e\u96c6\u4e0a\u7684\u8868\u73b0\u53ef\u80fd\u5b58\u5728\u5929\u7136\u4f18\u52bf\u3002\u8bc4\u4f30\u7ed3\u679c\u9700\u7ed3\u5408\u6a21\u578b\u80cc\u666f\u8fdb\u884c\u89e3\u8bfb\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u88c1\u5224\u6a21\u578b\u4f9d\u8d56<\/strong>:\n<ul>\n<li><code>token<\/code>\u00a0\u6548\u7387\u7684\u8ba1\u7b97\u4f9d\u8d56\u88c1\u5224\u6a21\u578b\uff08Judge Model\uff09\u51c6\u786e\u5224\u65ad\u63a8\u7406\u6b65\u9aa4\u7684\u6b63\u786e\u6027\u3002\u6b63\u5982\u00a0<code>ProcessBench<\/code>\u00a0<sup><a href=\"https:\/\/markdown.tchepai.com\/#footnotes-def-4\">4<\/a><\/sup>\u7814\u7a76\u6240\u6307\u51fa\u7684\uff0c\u8fd9\u5bf9\u73b0\u6709\u6a21\u578b\u800c\u8a00\u662f\u4e00\u9879\u6311\u6218\u6027\u4efb\u52a1\uff0c\u901a\u5e38\u9700\u8981\u80fd\u529b\u5f88\u5f3a\u7684\u6a21\u578b\u624d\u80fd\u80dc\u4efb\u3002<\/li>\n<li>\u88c1\u5224\u6a21\u578b\u7684\u8bef\u5224\u4f1a\u76f4\u63a5\u5f71\u54cd\u00a0<code>token<\/code>\u00a0\u6548\u7387\u6307\u6807\u7684\u51c6\u786e\u6027\uff0c\u56e0\u6b64\u9009\u62e9\u5408\u9002\u7684\u88c1\u5224\u6a21\u578b\u81f3\u5173\u91cd\u8981\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<p>\u603b\u800c\u8a00\u4e4b\uff0c<code>EvalThink<\/code>\u00a0\u63d0\u4f9b\u4e86\u4e00\u5957\u91cf\u5316\u8bc4\u4f30 LLM \u601d\u8003\u6548\u7387\u7684\u6846\u67b6\u548c\u6307\u6807\uff0c\u63ed\u793a\u4e86\u4e0d\u540c\u6a21\u578b\u5728\u51c6\u786e\u6027\u3001<code>token<\/code>\u00a0\u6d88\u8017\u548c\u601d\u8003\u6df1\u5ea6\u4e4b\u95f4\u7684\u6743\u8861\u3002\u8fd9\u4e9b\u53d1\u73b0\u5bf9\u4e8e\u6307\u5bfc\u6a21\u578b\u8bad\u7ec3\uff08\u5982 <a href=\"https:\/\/www.kdjingpai.com\/ja\/grpo-ruhezaishi\/\">GRPO<\/a> \u548c SFT\uff09\uff0c\u5f00\u53d1\u51fa\u66f4\u9ad8\u6548\u3001\u80fd\u6839\u636e\u95ee\u9898\u96be\u5ea6\u81ea\u9002\u5e94\u8c03\u6574\u601d\u8003\u6df1\u5ea6\u7684\u4e0b\u4e00\u4ee3\u6a21\u578b\uff0c\u5177\u6709\u91cd\u8981\u7684\u53c2\u8003\u4ef7\u503c\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u7684\u53d1\u5c55\u65e5\u65b0\u6708\u5f02\uff0c\u5176\u63a8\u7406\u80fd\u529b\u5df2\u6210\u4e3a\u8861\u91cf\u5176\u667a\u80fd\u6c34\u5e73\u7684\u5173\u952e\u6307\u6807\u3002\u7279\u522b\u662f\u5177\u5907\u957f\u63a8\u7406\u80fd\u529b\u7684\u6a21\u578b\uff0c\u4f8b\u5982 OpenAI \u7684\u00a0o1\u3001DeepSeek-R1\u3001QwQ-32B\u00a0\u548c\u00a0Kimi K1.5\u00a0\u7b49\uff0c\u5b83\u4eec\u901a\u8fc7\u6a21\u62df\u4eba\u7c7b\u6df1\u5ea6\u601d\u8003\u8fc7\u7a0b\u6765\u89e3\u51b3\u590d&#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[34],"tags":[],"class_list":["post-29685","post","type-post","status-publish","format-standard","hentry","category-knowledge"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/posts\/29685","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/comments?post=29685"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/posts\/29685\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/media?parent=29685"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/categories?post=29685"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/tags?post=29685"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}