{"id":51133,"date":"2025-08-28T05:46:26","date_gmt":"2025-08-27T21:46:26","guid":{"rendered":"https:\/\/www.kdjingpai.com\/?p=51133"},"modified":"2025-08-28T05:46:26","modified_gmt":"2025-08-27T21:46:26","slug":"qwen3-finetuning-playground","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/de\/qwen3-finetuning-playground\/","title":{"rendered":"Qwen3-FineTuning-Playground\uff1a\u4e00\u4e2a\u4e0a\u624b\u5373\u7528\u7684Qwen3\u5927\u6a21\u578b\u5fae\u8c03\u4ee3\u7801\u5e93"},"content":{"rendered":"<p>Qwen3-FineTuning-Playground\u662f\u4e00\u4e2a\u5f00\u6e90\u9879\u76ee\uff0c\u5b83\u63d0\u4f9b\u4e86\u4e00\u5957\u5b8c\u6574\u7684\u4ee3\u7801\u5e93\uff0c\u4e13\u95e8\u7528\u4e8e\u5bf9Qwen3\u7cfb\u5217\u5927\u8bed\u8a00\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\u3002\u8fd9\u4e2a\u9879\u76ee\u7684\u57fa\u7840\u662f\u63d0\u4f9b\u6e05\u6670\u3001\u4e13\u4e1a\u4e14\u5bb9\u6613\u6269\u5c55\u7684\u5fae\u8c03\u4ee3\u7801\u793a\u4f8b\uff0c\u8ba9\u5f00\u53d1\u8005\u548c\u7814\u7a76\u4eba\u5458\u53ef\u4ee5\u65b9\u4fbf\u5730\u5b9e\u8df5\u5404\u79cd\u4e3b\u6d41\u7684\u5fae\u8c03\u6280\u672f\u3002\u9879\u76ee\u4ee3\u7801\u7ed3\u6784\u6e05\u6670\uff0c\u5c06\u4e0d\u540c\u529f\u80fd\u6a21\u5757\u5316\uff0c\u4f8b\u5982\u76d1\u7763\u5fae\u8c03\u3001\u5f3a\u5316\u5b66\u4e60\u3001\u77e5\u8bc6\u84b8\u998f\u3001\u63a8\u7406\u7b49\u90fd\u5b58\u653e\u5728\u72ec\u7acb\u7684\u76ee\u5f55\u4e2d\u3002\u6240\u6709\u7684\u8bad\u7ec3\u548c\u63a8\u7406\u811a\u672c\u90fd\u652f\u6301\u901a\u8fc7\u547d\u4ee4\u884c\u4f20\u9012\u53c2\u6570\u6765\u914d\u7f6e\uff0c\u8fd9\u610f\u5473\u7740\u4f7f\u7528\u8005\u4e0d\u9700\u8981\u4fee\u6539\u6e90\u4ee3\u7801\u5c31\u53ef\u4ee5\u8fd0\u884c\u4e0d\u540c\u7684\u5b9e\u9a8c\u3002\u6b64\u5916\uff0c\u9879\u76ee\u8fd8\u63d0\u4f9b\u4e86\u591a\u4e2a\u8be6\u7ec6\u7684\u7aef\u5230\u7aef\u6559\u7a0b\uff0c\u5373\u4fbf\u662f\u521d\u5b66\u8005\uff0c\u4e5f\u80fd\u6309\u7167\u6587\u6863\u8bf4\u660e\uff0c\u4e00\u6b65\u6b65\u5b8c\u6210\u4ece\u73af\u5883\u914d\u7f6e\u3001\u6570\u636e\u51c6\u5907\u5230\u6a21\u578b\u8bad\u7ec3\u548c\u63a8\u7406\u7684\u5168\u8fc7\u7a0b\u3002<\/p>\n<h2>\u529f\u80fd\u5217\u8868<\/h2>\n<ul>\n<li><strong>\u591a\u79cd\u76d1\u7763\u5fae\u8c03\u65b9\u6848 (SFT)<\/strong>: \u652f\u6301\u5bf9\u6a21\u578b\u8fdb\u884c\u5168\u91cf\u53c2\u6570\u5fae\u8c03\uff0c\u4e5f\u652f\u6301\u4f7f\u7528LoRA\uff08\u4f4e\u79e9\u9002\u5e94\uff09\u8fd9\u7c7b\u6280\u672f\u8fdb\u884c\u9ad8\u6548\u5fae\u8c03\uff0c\u4ee5\u9002\u5e94\u4e0d\u540c\u7684\u786c\u4ef6\u8d44\u6e90\u548c\u8bad\u7ec3\u9700\u6c42\u3002<\/li>\n<li><strong>\u5f3a\u5316\u5b66\u4e60\u5bf9\u9f50 (RL)<\/strong>: \u96c6\u6210\u4e86\u591a\u79cd\u57fa\u4e8e\u4eba\u7c7b\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\uff0c\u7528\u4e8e\u63d0\u5347\u6a21\u578b\u7684\u5bf9\u8bdd\u8d28\u91cf\u548c\u9075\u5faa\u6307\u4ee4\u7684\u80fd\u529b\u3002\n<ul>\n<li><strong>PPO\u7b97\u6cd5<\/strong>: \u5b9e\u73b0\u4e86\u7ecf\u5178\u7684\u8fd1\u7aef\u7b56\u7565\u4f18\u5316\u7b97\u6cd5\uff0c\u901a\u8fc7\u5956\u52b1\u6a21\u578b\uff08RM\uff09\u6765\u6307\u5bfc\u6a21\u578b\u5b66\u4e60\u3002<\/li>\n<li><strong>ORPO\u7b97\u6cd5<\/strong>: \u63d0\u4f9b\u4e86\u4e00\u79cd\u65e0\u9700\u989d\u5916\u5956\u52b1\u6a21\u578b\u7684\u9ad8\u6548\u504f\u597d\u5bf9\u9f50\u7b97\u6cd5\uff0c\u7b80\u5316\u4e86\u8bad\u7ec3\u6d41\u7a0b\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u8bad\u540e\u4f18\u5316\u6280\u672f<\/strong>:\n<ul>\n<li><strong>\u77e5\u8bc6\u84b8\u998f<\/strong>: \u652f\u6301\u5c06\u4e00\u4e2a\u66f4\u5927\u3001\u80fd\u529b\u66f4\u5f3a\u7684\u6559\u5e08\u6a21\u578b\uff08\u5982Qwen3-4B\uff09\u7684\u77e5\u8bc6\u8fc1\u79fb\u5230\u4e00\u4e2a\u66f4\u5c0f\u7684\u5b66\u751f\u6a21\u578b\uff08\u5982Qwen3-1.7B\uff09\uff0c\u4ee5\u83b7\u5f97\u4e00\u4e2a\u8f7b\u91cf\u4f46\u6027\u80fd\u4f9d\u7136\u5f3a\u5927\u7684\u6a21\u578b\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u6a21\u5757\u5316\u7684\u4ee3\u7801\u7ed3\u6784<\/strong>: \u9879\u76ee\u5c06\u4ee3\u7801\u6309\u7167\u529f\u80fd\u6e05\u6670\u5730\u5212\u5206\u5230\u4e0d\u540c\u76ee\u5f55\uff0c\u4f8b\u5982<code>Supervised_FineTuning<\/code>\u3001<code>RL_FineTuning<\/code>\u7b49\uff0c\u4f7f\u4ee3\u7801\u6613\u4e8e\u7406\u89e3\u548c\u7ef4\u62a4\u3002<\/li>\n<li><strong>\u53c2\u6570\u5316\u8fd0\u884c\u811a\u672c<\/strong>: \u6240\u6709\u811a\u672c\u90fd\u652f\u6301\u901a\u8fc7\u547d\u4ee4\u884c\u53c2\u6570\u8fdb\u884c\u914d\u7f6e\uff0c\u7528\u6237\u53ef\u4ee5\u7075\u6d3b\u5730\u8c03\u6574\u6a21\u578b\u8def\u5f84\u3001\u6570\u636e\u96c6\u3001\u8f93\u51fa\u76ee\u5f55\u7b49\u53c2\u6570\uff0c\u800c\u65e0\u9700\u4fee\u6539\u4ee3\u7801\u3002<\/li>\n<li><strong>\u7aef\u5230\u7aef\u793a\u4f8b\u6559\u7a0b<\/strong>: \u5728<code>example\/<\/code>\u76ee\u5f55\u4e0b\u63d0\u4f9b\u4e86\u591a\u4e2a\u5b8c\u6574\u7684\u5b9e\u6218\u6559\u7a0b\uff0c\u8986\u76d6\u4e86\u4eceSFT\u5230PPO\u7684\u5b8c\u6574\u6d41\u7a0b\u3001ORPO\u7684\u5355\u6b65\u5bf9\u9f50\u3001\u77e5\u8bc6\u84b8\u998f\u4ee5\u53ca\u7279\u5b9a\u9886\u57df\u7684\u5e94\u7528\uff08\u5982\u5fc3\u7406\u5b66\u591a\u8f6e\u5bf9\u8bdd\u5fae\u8c03\uff09\u3002<\/li>\n<\/ul>\n<h2>\u4f7f\u7528\u5e2e\u52a9<\/h2>\n<p>\u8fd9\u4e2a\u4ee3\u7801\u5e93\u63d0\u4f9b\u4e86\u4e00\u4e2a\u5b8c\u6574\u7684\u6d41\u7a0b\uff0c\u5e2e\u52a9\u4f60\u5feb\u901f\u4e0a\u624b\u5bf9Qwen3\u5927\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\u3002\u4e0b\u9762\u5c06\u4ee5\u4e00\u4e2aSFT-LoRA\u5fae\u8c03\u4e3a\u4f8b\uff0c\u8be6\u7ec6\u4ecb\u7ecd\u64cd\u4f5c\u6b65\u9aa4\u3002<\/p>\n<h3>1. \u51c6\u5907\u5de5\u4f5c\uff1a\u514b\u9686\u9879\u76ee\u548c\u914d\u7f6e\u73af\u5883<\/h3>\n<p>\u9996\u5148\uff0c\u4f60\u9700\u8981\u5c06\u9879\u76ee\u4ee3\u7801\u4eceGitHub\u514b\u9686\u5230\u4f60\u7684\u672c\u5730\u8ba1\u7b97\u673a\uff0c\u5e76\u8fdb\u5165\u9879\u76ee\u76ee\u5f55\u3002<\/p>\n<pre><code>git clone https:\/\/github.com\/Muziqinshan\/Qwen3-FineTuning-Playground.git\r\ncd Qwen3-FineTuning-Playground\r\n<\/code><\/pre>\n<p>\u63a5\u4e0b\u6765\u662f\u914d\u7f6ePython\u73af\u5883\u3002\u4e3a\u4e86\u907f\u514d\u4e0e\u4f60\u7cfb\u7edf\u4e2d\u5df2\u6709\u7684\u5e93\u4ea7\u751f\u51b2\u7a81\uff0c\u5f3a\u70c8\u5efa\u8bae\u4f7f\u7528<code>conda<\/code>\u521b\u5efa\u4e00\u4e2a\u5168\u65b0\u7684\u72ec\u7acb\u73af\u5883\u3002\u8fd9\u4e2a\u9879\u76ee\u63a8\u8350\u4f7f\u7528Python 3.10\u7248\u672c\u3002<\/p>\n<pre><code># \u521b\u5efa\u4e00\u4e2a\u540d\u4e3a qwen3_ft \u7684\u65b0\u73af\u5883\r\nconda create -n qwen3_ft python=3.10\r\n# \u6fc0\u6d3b\u8fd9\u4e2a\u65b0\u521b\u5efa\u7684\u73af\u5883\r\nconda activate qwen3_ft\r\n<\/code><\/pre>\n<p>\u73af\u5883\u6fc0\u6d3b\u540e\uff0c\u4f60\u9700\u8981\u5b89\u88c5\u9879\u76ee\u6240\u9700\u7684\u6240\u6709\u4f9d\u8d56\u5e93\u3002\u9879\u76ee\u6839\u76ee\u5f55\u4e0b\u7684<code>requirements.txt<\/code>\u6587\u4ef6\u5df2\u7ecf\u5217\u51fa\u4e86\u6240\u6709\u5fc5\u9700\u7684\u5e93\u3002\u8fd0\u884c\u4ee5\u4e0b\u547d\u4ee4\u8fdb\u884c\u5b89\u88c5\uff1a<\/p>\n<pre><code>pip install -r requirements.txt\r\n<\/code><\/pre>\n<h3>2. \u4e0b\u8f7d\u6a21\u578b\u548c\u51c6\u5907\u6570\u636e<\/h3>\n<p>\u5728\u5fae\u8c03\u4e4b\u524d\uff0c\u4f60\u9700\u8981\u51c6\u5907\u597d\u57fa\u7840\u6a21\u578b\u548c\u7528\u4e8e\u8bad\u7ec3\u7684\u6570\u636e\u96c6\u3002<\/p>\n<p><strong>\u4e0b\u8f7d\u6a21\u578b<\/strong><br \/>\n\u672c\u9879\u76ee\u63a8\u8350\u4f7f\u7528<code>modelscope<\/code>\u5e93\u4ece\u201c\u9b54\u642d\u201d\u793e\u533a\u4e0b\u8f7d\u9884\u8bad\u7ec3\u597d\u7684Qwen3\u6a21\u578b\u3002<code>modelscope<\/code>\u5e93\u5728\u6267\u884c\u4e0a\u4e00\u6b65<code>pip install<\/code>\u547d\u4ee4\u65f6\u5df2\u7ecf\u88ab\u5b89\u88c5\u3002<\/p>\n<p>\u8fd0\u884c\u4ee5\u4e0b\u547d\u4ee4\u6765\u4e0b\u8f7d\u672c\u9879\u76ee\u793a\u4f8b\u4e2d\u9700\u8981\u7528\u5230\u7684\u4e24\u4e2a\u57fa\u7840\u6a21\u578b\uff1a<\/p>\n<pre><code># \u4e0b\u8f7dQwen3-1.7B\u6a21\u578b\uff0c\u5b83\u5c06\u4e3b\u8981\u7528\u4e8eSFT\u3001ORPO\u548cPPO\u7b49\u5fae\u8c03\u4efb\u52a1\r\nmodelscope download --model Qwen\/Qwen3-1.7B --local_dir .\/Qwen3\/Qwen3-1.7B\r\n# \u4e0b\u8f7dQwen3-4B\u6a21\u578b\uff0c\u5b83\u4e3b\u8981\u7528\u4f5c\u77e5\u8bc6\u84b8\u998f\u4efb\u52a1\u4e2d\u7684\u6559\u5e08\u6a21\u578b\r\nmodelscope download --model Qwen\/Qwen3-4B --local_dir .\/Qwen3\/Qwen3-4B\r\n<\/code><\/pre>\n<p>\u547d\u4ee4\u6267\u884c\u540e\uff0c\u6a21\u578b\u6587\u4ef6\u4f1a\u81ea\u52a8\u4e0b\u8f7d\u5e76\u4fdd\u5b58\u5230\u9879\u76ee\u6839\u76ee\u5f55\u4e0b\u7684<code>.\/Qwen3\/Qwen3-1.7B<\/code>\u548c<code>.\/Qwen3\/Qwen3-4B<\/code>\u76ee\u5f55\u4e2d\u3002<\/p>\n<p><strong>\u51c6\u5907\u6570\u636e<\/strong><br \/>\n\u672c\u9879\u76ee\u4f7f\u7528\u7684\u6570\u636e\u6709\u7279\u5b9a\u7684JSON\u683c\u5f0f\u3002\u5728<code>data\/<\/code>\u76ee\u5f55\u4e0b\uff0c\u63d0\u4f9b\u4e86\u4e00\u4e2a\u540d\u4e3a<code>dirty_chinese_dpo.json<\/code>\u7684\u793a\u4f8b\u6587\u4ef6\uff0c\u4f60\u53ef\u4ee5\u53c2\u8003\u5b83\u7684\u683c\u5f0f\u6765\u51c6\u5907\u4f60\u81ea\u5df1\u7684\u6570\u636e\u96c6\u3002<\/p>\n<h3>3. \u5f00\u59cbSFT-LoRA\u5fae\u8c03<\/h3>\n<p>\u4e00\u5207\u51c6\u5907\u5c31\u7eea\u540e\uff0c\u5c31\u53ef\u4ee5\u5f00\u59cb\u8bad\u7ec3\u4e86\u3002\u4e0b\u9762\u7684\u547d\u4ee4\u5c06\u542f\u52a8\u4e00\u4e2a\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\u4efb\u52a1\uff0c\u5e76\u4f7f\u7528LoRA\u6280\u672f\u6765\u63d0\u5347\u6548\u7387\u3002<\/p>\n<pre><code>python Supervised_FineTuning\/train_sft_dirty.py \\\r\n--model_path .\/Qwen3\/Qwen3-1.7B \\\r\n--dataset_path data\/dirty_chinese_dpo.json \\\r\n--sft_adapter_output_dir .\/output\/sft_adapter_demo\r\n<\/code><\/pre>\n<ul>\n<li><code>--model_path<\/code>: \u6307\u5b9a\u4e86\u6211\u4eec\u521a\u521a\u4e0b\u8f7d\u7684\u57fa\u7840\u6a21\u578b\u7684\u8def\u5f84\u3002<\/li>\n<li><code>--dataset_path<\/code>: \u6307\u5b9a\u4e86\u7528\u4e8e\u8bad\u7ec3\u7684\u6570\u636e\u96c6\u6587\u4ef6\u3002<\/li>\n<li><code>--sft_adapter_output_dir<\/code>: \u6307\u5b9a\u4e86\u8bad\u7ec3\u5b8c\u6210\u540e\uff0cLoRA\u9002\u914d\u5668\u6743\u91cd\u4fdd\u5b58\u7684\u76ee\u5f55\u3002<\/li>\n<\/ul>\n<p>\u8bad\u7ec3\u8fc7\u7a0b\u4f1a\u6839\u636e\u4f60\u7684\u786c\u4ef6\u6027\u80fd\u6301\u7eed\u4e00\u6bb5\u65f6\u95f4\u3002\u8bad\u7ec3\u7ed3\u675f\u540e\uff0c\u4f60\u4f1a\u5728<code>.\/output\/sft_adapter_demo<\/code>\u76ee\u5f55\u4e0b\u627e\u5230\u751f\u6210\u7684\u6a21\u578b\u9002\u914d\u5668\u6587\u4ef6\u3002<\/p>\n<h3>4. \u4f7f\u7528\u5fae\u8c03\u540e\u7684\u6a21\u578b\u8fdb\u884c\u63a8\u7406<\/h3>\n<p>\u6a21\u578b\u8bad\u7ec3\u5b8c\u6210\u540e\uff0c\u6700\u91cd\u8981\u7684\u4e00\u6b65\u5c31\u662f\u9a8c\u8bc1\u5b83\u7684\u6548\u679c\u3002\u4f60\u53ef\u4ee5\u8fd0\u884c\u4e0b\u9762\u7684\u63a8\u7406\u811a\u672c\uff0c\u4e0e\u4f60\u521a\u521a\u5fae\u8c03\u597d\u7684\u6a21\u578b\u8fdb\u884c\u4e92\u52a8\u804a\u5929\u3002<\/p>\n<pre><code>python inference\/inference_dirty_sft.py \\\r\n--model_path .\/Qwen3\/Qwen3-1.7B \\\r\n--adapter_path .\/output\/sft_adapter_demo \\\r\n--mode interactive\r\n<\/code><\/pre>\n<ul>\n<li><code>--model_path<\/code>: \u540c\u6837\u662f\u57fa\u7840\u6a21\u578b\u7684\u8def\u5f84\u3002<\/li>\n<li><code>--adapter_path<\/code>: \u6307\u5411\u6211\u4eec\u4e0a\u4e00\u6b65\u8bad\u7ec3\u597d\u7684LoRA\u9002\u914d\u5668\u6240\u5728\u7684\u76ee\u5f55\u3002<\/li>\n<li><code>--mode interactive<\/code>: \u8868\u793a\u542f\u52a8\u4ea4\u4e92\u5f0f\u804a\u5929\u6a21\u5f0f\uff0c\u4f60\u53ef\u4ee5\u5728\u547d\u4ee4\u884c\u4e2d\u76f4\u63a5\u8f93\u5165\u95ee\u9898\u4e0e\u6a21\u578b\u5bf9\u8bdd\u3002<\/li>\n<\/ul>\n<p>\u73b0\u5728\uff0c\u4f60\u5c31\u53ef\u4ee5\u6d4b\u8bd5\u6a21\u578b\u5728\u7279\u5b9a\u4efb\u52a1\u4e0a\u7684\u8868\u73b0\u4e86\u3002\u5982\u679c\u4f60\u60f3\u5c1d\u8bd5\u5176\u4ed6\u5fae\u8c03\u65b9\u6cd5\uff0c\u6bd4\u5982ORPO\u6216\u77e5\u8bc6\u84b8\u998f\uff0c\u53ef\u4ee5\u53c2\u8003<code>example\/<\/code>\u76ee\u5f55\u4e0b\u7684\u8be6\u7ec6\u6559\u7a0b\u6587\u6863\uff0c\u6b65\u9aa4\u4e0e\u6b64\u7c7b\u4f3c\u3002<\/p>\n<h2>\u5e94\u7528\u573a\u666f<\/h2>\n<ol>\n<li><strong>\u7279\u5b9a\u9886\u57df\u667a\u80fd\u5ba2\u670d<\/strong><br \/>\n\u53ef\u4ee5\u5c06\u901a\u7528\u7684Qwen3\u6a21\u578b\u5728\u7279\u5b9a\u884c\u4e1a\u7684\u77e5\u8bc6\u5e93\uff08\u4f8b\u5982\u91d1\u878d\u3001\u533b\u7597\u3001\u6cd5\u5f8b\uff09\u4e0a\u8fdb\u884c\u5fae\u8c03\uff0c\u4f7f\u5176\u6210\u4e3a\u80fd\u591f\u51c6\u786e\u56de\u7b54\u4e13\u4e1a\u95ee\u9898\u7684\u667a\u80fd\u5ba2\u670d\u673a\u5668\u4eba\u3002\u9879\u76ee\u4e2d\u63d0\u4f9b\u7684SFT\u6559\u7a0b\u662f\u5b9e\u73b0\u8fd9\u4e00\u573a\u666f\u7684\u7406\u60f3\u8d77\u70b9\u3002<\/li>\n<li><strong>\u4e2a\u6027\u5316\u5185\u5bb9\u521b\u4f5c\u52a9\u624b<\/strong><br \/>\n\u901a\u8fc7\u4f7f\u7528\u7279\u5b9a\u98ce\u683c\u7684\u6587\u672c\u6570\u636e\uff08\u5982\u67d0\u4e2a\u4f5c\u5bb6\u7684\u4f5c\u54c1\u3001\u7279\u5b9a\u98ce\u683c\u7684\u8425\u9500\u6587\u6848\uff09\u5bf9\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\uff0c\u53ef\u4ee5\u521b\u9020\u51fa\u4e00\u4e2a\u80fd\u591f\u6a21\u4eff\u8be5\u98ce\u683c\u8fdb\u884c\u5199\u4f5c\u7684\u52a9\u624b\uff0c\u7528\u4e8e\u8f85\u52a9\u5185\u5bb9\u521b\u4f5c\u3001\u8bd7\u6b4c\u5199\u4f5c\u6216\u5e7f\u544a\u6587\u6848\u751f\u6210\u3002<\/li>\n<li><strong>\u6a21\u578b\u8f7b\u91cf\u5316\u4e0e\u79c1\u6709\u5316\u90e8\u7f72<\/strong><br \/>\n\u5bf9\u4e8e\u7b97\u529b\u8d44\u6e90\u6709\u9650\u7684\u573a\u666f\uff0c\u53ef\u4ee5\u4f7f\u7528\u77e5\u8bc6\u84b8\u998f\u529f\u80fd\uff0c\u5c06\u4e00\u4e2a\u5927\u578b\u6a21\u578b\uff08\u5982Qwen3-4B\uff09\u7684\u80fd\u529b\u8fc1\u79fb\u5230\u4e00\u4e2a\u5c0f\u578b\u6a21\u578b\uff08\u5982Qwen3-1.7B\uff09\u4e0a\u3002\u8fd9\u6837\u65e2\u80fd\u4fdd\u7559\u5927\u90e8\u5206\u6027\u80fd\uff0c\u53c8\u80fd\u663e\u8457\u964d\u4f4e\u63a8\u7406\u6210\u672c\uff0c\u4fbf\u4e8e\u5728\u4e2a\u4eba\u8bbe\u5907\u6216\u8fb9\u7f18\u8ba1\u7b97\u8bbe\u5907\u4e0a\u8fdb\u884c\u79c1\u6709\u5316\u90e8\u7f72\u3002<\/li>\n<li><strong>\u63d0\u5347\u6a21\u578b\u5bf9\u8bdd\u7684\u5b89\u5168\u6027\u4e0e\u4e00\u81f4\u6027<\/strong><br \/>\n\u53ef\u4ee5\u4f7f\u7528\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\uff08\u5982PPO\u6216ORPO\uff09\uff0c\u901a\u8fc7\u4eba\u7c7b\u504f\u597d\u6570\u636e\u5bf9\u6a21\u578b\u8fdb\u884c\u5bf9\u9f50\uff0c\u4f7f\u5176\u8f93\u51fa\u7684\u56de\u7b54\u66f4\u7b26\u5408\u4eba\u7c7b\u7684\u4ef7\u503c\u89c2\uff0c\u51cf\u5c11\u6709\u5bb3\u6216\u4e0d\u5f53\u5185\u5bb9\u7684\u751f\u6210\uff0c\u63d0\u5347\u5bf9\u8bdd\u7684\u53ef\u9760\u6027\u548c\u5b89\u5168\u6027\u3002<\/li>\n<\/ol>\n<h2>QA<\/h2>\n<ol>\n<li><strong>\u8fd9\u4e2a\u9879\u76ee\u4e3b\u8981\u652f\u6301\u54ea\u4e9bQwen3\u6a21\u578b\uff1f<\/strong><br \/>\n\u8be5\u9879\u76ee\u4e3b\u8981\u57fa\u4e8eQwen3\u7cfb\u5217\u6a21\u578b\u8fdb\u884c\u5f00\u53d1\u548c\u6d4b\u8bd5\uff0c\u793a\u4f8b\u4ee3\u7801\u4e2d\u76f4\u63a5\u63d0\u4f9b\u4e86Qwen3-1.7B\u548cQwen3-4B\u7684\u6a21\u578b\u4e0b\u8f7d\u548c\u4f7f\u7528\u65b9\u6cd5\u3002\u7406\u8bba\u4e0a\uff0c\u4ee3\u7801\u7ed3\u6784\u4e5f\u517c\u5bb9Qwen3\u7cfb\u5217\u7684\u5176\u4ed6\u5c3a\u5bf8\u6a21\u578b\u3002<\/li>\n<li><strong>\u8fdb\u884c\u5fae\u8c03\u9700\u8981\u4ec0\u4e48\u6837\u7684\u786c\u4ef6\u914d\u7f6e\uff1f<\/strong><br \/>\n\u786c\u4ef6\u9700\u6c42\u53d6\u51b3\u4e8e\u4f60\u9009\u62e9\u7684\u5fae\u8c03\u65b9\u6cd5\u548c\u6a21\u578b\u5927\u5c0f\u3002\u5bf9\u4e8eSFT-LoRA\u5fae\u8c03Qwen3-1.7B\u8fd9\u6837\u7684\u5c0f\u6a21\u578b\uff0c\u4e00\u5f20\u6d88\u8d39\u7ea7\u7684\u663e\u5361\uff08\u5982NVIDIA 3090\u62164090\uff09\u901a\u5e38\u5c31\u8db3\u591f\u4e86\u3002\u4f46\u5982\u679c\u8981\u8fdb\u884c\u5168\u91cf\u5fae\u8c03\u6216\u8005\u8bad\u7ec3\u66f4\u5927\u7684\u6a21\u578b\uff0c\u5219\u9700\u8981\u66f4\u591a\u7684\u663e\u5b58\u548c\u8ba1\u7b97\u8d44\u6e90\u3002<\/li>\n<li><strong>ORPO\u548cPPO\u5fae\u8c03\u6709\u4ec0\u4e48\u533a\u522b\uff1f<\/strong><br \/>\nPPO\u662f\u4e00\u79cd\u7ecf\u5178\u7684\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\uff0c\u5b83\u9700\u8981\u4e00\u4e2a\u72ec\u7acb\u7684\u3001\u9884\u5148\u8bad\u7ec3\u597d\u7684\u5956\u52b1\u6a21\u578b\uff08Reward Model\uff09\u6765\u4e3a\u6a21\u578b\u7684\u8f93\u51fa\u6253\u5206\uff0c\u8bad\u7ec3\u6d41\u7a0b\u76f8\u5bf9\u590d\u6742\u3002\u800cORPO\u662f\u4e00\u79cd\u66f4\u65b0\u7684\u7b97\u6cd5\uff0c\u5b83\u4e0d\u9700\u8981\u989d\u5916\u7684\u5956\u52b1\u6a21\u578b\uff0c\u76f4\u63a5\u901a\u8fc7\u504f\u597d\u6570\u636e\u5bf9\u6a21\u578b\u8fdb\u884c\u4f18\u5316\uff0c\u6d41\u7a0b\u66f4\u7b80\u5355\uff0c\u6548\u7387\u66f4\u9ad8\u3002<\/li>\n<li><strong>\u6211\u53ef\u4ee5\u4f7f\u7528\u81ea\u5df1\u7684\u6570\u636e\u96c6\u5417\uff1f<\/strong><br \/>\n\u5b8c\u5168\u53ef\u4ee5\u3002\u4f60\u53ea\u9700\u8981\u5c06\u4f60\u7684\u6570\u636e\u96c6\u6574\u7406\u6210\u9879\u76ee<code>data\/<\/code>\u76ee\u5f55\u4e0b\u793a\u4f8b\u6587\u4ef6\u90a3\u6837\u7684JSON\u683c\u5f0f\uff0c\u7136\u540e\u5728\u8fd0\u884c\u8bad\u7ec3\u811a\u672c\u65f6\uff0c\u901a\u8fc7<code>--dataset_path<\/code>\u53c2\u6570\u6307\u5b9a\u4f60\u7684\u6587\u4ef6\u8def\u5f84\u5373\u53ef\u3002<\/li>\n<\/ol>\n","protected":false},"excerpt":{"rendered":"<p>Qwen3-FineTuning-Playground\u662f\u4e00\u4e2a\u5f00\u6e90\u9879\u76ee\uff0c\u5b83\u63d0\u4f9b\u4e86\u4e00\u5957\u5b8c\u6574\u7684\u4ee3\u7801\u5e93\uff0c\u4e13\u95e8\u7528\u4e8e\u5bf9Qwen3\u7cfb\u5217\u5927\u8bed\u8a00\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\u3002\u8fd9\u4e2a\u9879\u76ee\u7684\u57fa\u7840\u662f\u63d0\u4f9b\u6e05\u6670\u3001\u4e13\u4e1a\u4e14\u5bb9\u6613\u6269\u5c55\u7684\u5fae\u8c03\u4ee3\u7801\u793a\u4f8b\uff0c\u8ba9\u5f00\u53d1\u8005\u548c\u7814\u7a76\u4eba\u5458\u53ef\u4ee5\u65b9\u4fbf\u5730\u5b9e\u8df5\u5404\u79cd\u4e3b\u6d41\u7684\u5fae&#8230;<\/p>\n","protected":false},"author":1,"featured_media":32782,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[20,483],"tags":[230,365],"class_list":["post-51133","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-tool","category-fine-tuning","tag-aikaiyuanxiangmu","tag-damoxingweidiao"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/posts\/51133","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=51133"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/posts\/51133\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/media\/32782"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/media?parent=51133"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/categories?post=51133"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/tags?post=51133"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}