{"id":29379,"date":"2025-03-26T01:24:20","date_gmt":"2025-03-25T17:24:20","guid":{"rendered":"https:\/\/www.aisharenet.com\/?p=29379"},"modified":"2025-03-26T01:24:20","modified_gmt":"2025-03-25T17:24:20","slug":"bonsai","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/en\/bonsai\/","title":{"rendered":"Bonsai\uff1a\u9002\u5408\u8fb9\u7f18\u8bbe\u5907\u8fd0\u884c\u7684\u4e09\u503c\u6743\u91cd\u8bed\u8a00\u6a21\u578b"},"content":{"rendered":"<p>Bonsai \u662f deepgrove-ai \u5f00\u53d1\u7684\u4e00\u4e2a\u5f00\u6e90\u8bed\u8a00\u6a21\u578b\uff0c\u53c2\u6570\u89c4\u6a21\u4e3a 5 \u4ebf\uff0c\u91c7\u7528\u4e09\u503c\u6743\u91cd\uff08ternary weights\uff09\u6280\u672f\u3002\u5b83\u57fa\u4e8e Llama \u67b6\u6784\u548c <a href=\"https:\/\/www.kdjingpai.com\/en\/le-chat-mistral\/\">Mistral<\/a> \u5206\u8bcd\u5668\u8bbe\u8ba1\uff0c\u7ebf\u6027\u5c42\u7ecf\u8fc7\u8c03\u6574\u4ee5\u652f\u6301\u4e09\u503c\u6743\u91cd\u3002\u6a21\u578b\u4e3b\u8981\u4f7f\u7528 DCLM-Pro \u548c Fineweb-Edu \u6570\u636e\u96c6\u8bad\u7ec3\uff0c\u603b\u8ba1\u4e0d\u5230 50 \u4ebf\u4e2a\u6807\u8bb0\u3002\u5c3d\u7ba1\u8bad\u7ec3\u6570\u636e\u5c11\uff0cBonsai \u6027\u80fd\u4f9d\u7136\u51fa\u8272\uff0c\u662f\u9996\u6279\u8fbe\u5230\u7ade\u4e89\u6c34\u5e73\u7684\u8f7b\u91cf\u4e09\u503c\u6a21\u578b\u4e4b\u4e00\u3002\u7528\u6237\u53ef\u4ee5\u901a\u8fc7 Huggingface Transformers \u5e93\u8c03\u7528\u5b83\u3002\u9879\u76ee\u4ee3\u7801\u5728 GitHub \u4e0a\u516c\u5f00\uff0c\u9002\u5408\u5f00\u53d1\u8005\u63a2\u7d22\u9ad8\u6548 AI \u6a21\u578b\u3002<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-29380\" title=\"Bonsai\uff1a\u8f7b\u91cf\u9ad8\u6548\u7684\u4e09\u503c\u6743\u91cd\u8bed\u8a00\u6a21\u578b-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/a7365472e2e8a29.png\" alt=\"Bonsai\uff1a\u8f7b\u91cf\u9ad8\u6548\u7684\u4e09\u503c\u6743\u91cd\u8bed\u8a00\u6a21\u578b-1\" width=\"499\" height=\"301\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/a7365472e2e8a29.png 499w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/03\/a7365472e2e8a29-18x12.png 18w\" sizes=\"auto, (max-width: 499px) 100vw, 499px\" \/><\/p>\n<p>&nbsp;<\/p>\n<h2>\u529f\u80fd\u5217\u8868<\/h2>\n<ul>\n<li><strong>\u8f7b\u91cf\u9ad8\u6548\u8fd0\u884c<\/strong>\uff1a\u91c7\u7528\u4e09\u503c\u6743\u91cd\u6280\u672f\uff0c\u6a21\u578b\u5c0f\u5de7\uff0c\u8fd0\u884c\u901f\u5ea6\u5feb\uff0c\u9002\u5408\u4f4e\u8d44\u6e90\u8bbe\u5907\u3002<\/li>\n<li><strong>\u751f\u6210\u81ea\u7136\u8bed\u8a00<\/strong>\uff1a\u652f\u6301\u751f\u6210\u6d41\u7545\u6587\u672c\uff0c\u53ef\u7528\u4e8e\u5bf9\u8bdd\u3001\u95ee\u7b54\u7b49\u4efb\u52a1\u3002<\/li>\n<li><strong>\u5f00\u6e90\u8bbf\u95ee<\/strong>\uff1a\u5728 GitHub \u4e0a\u63d0\u4f9b\u5b8c\u6574\u4ee3\u7801\uff0c\u5141\u8bb8\u7528\u6237\u4e0b\u8f7d\u3001\u4fee\u6539\u548c\u4f18\u5316\u3002<\/li>\n<li><strong>\u517c\u5bb9 Huggingface<\/strong>\uff1a\u65e0\u7f1d\u96c6\u6210\u5230 Transformers \u5e93\uff0c\u4fbf\u4e8e\u52a0\u8f7d\u548c\u90e8\u7f72\u3002<\/li>\n<li><strong>\u6027\u80fd\u4f18\u5f02<\/strong>\uff1a\u5728\u5c11\u91cf\u8bad\u7ec3\u6570\u636e\u4e0b\uff0c\u6027\u80fd\u53ef\u5ab2\u7f8e\u540c\u7ea7\u522b\u6a21\u578b\u3002<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<h2>\u4f7f\u7528\u5e2e\u52a9<\/h2>\n<h3>\u5b89\u88c5\u6d41\u7a0b<\/h3>\n<p>\u8981\u4f7f\u7528 Bonsai\uff0c\u9700\u8981\u5148\u642d\u5efa\u8fd0\u884c\u73af\u5883\u3002\u4ee5\u4e0b\u662f\u8be6\u7ec6\u6b65\u9aa4\uff1a<\/p>\n<ol>\n<li><strong>\u68c0\u67e5 Python \u73af\u5883<\/strong><br \/>\n\u786e\u4fdd\u7535\u8111\u5b89\u88c5\u4e86 Python 3.8 \u6216\u4ee5\u4e0a\u7248\u672c\u3002\u5728\u7ec8\u7aef\u8f93\u5165\uff1a<\/li>\n<\/ol>\n<pre><code>python --version\r\n<\/code><\/pre>\n<p>\u5982\u679c\u672a\u5b89\u88c5\uff0c\u53ef\u4ece\u00a0Python \u5b98\u7f51\u00a0\u4e0b\u8f7d\u3002<\/p>\n<ol start=\"2\">\n<li><strong>\u5b89\u88c5 Transformers \u5e93<\/strong><br \/>\nBonsai \u4f9d\u8d56 Huggingface \u7684 Transformers \u5e93\u3002\u5728\u7ec8\u7aef\u8fd0\u884c\uff1a<\/li>\n<\/ol>\n<pre><code>pip install transformers\r\n<\/code><\/pre>\n<p>\u5b89\u88c5\u540e\uff0c\u7528\u00a0<code>pip show transformers<\/code>\u00a0\u786e\u8ba4\u7248\u672c\u3002<\/p>\n<ol start=\"3\">\n<li><strong>\u4e0b\u8f7d Bonsai \u6a21\u578b<\/strong><br \/>\n\u6a21\u578b\u6258\u7ba1\u5728 Huggingface \u4e0a\u3002\u63a8\u8350\u901a\u8fc7\u4ee3\u7801\u81ea\u52a8\u52a0\u8f7d\uff08\u89c1\u4e0b\u6587\uff09\uff0c\u4e5f\u53ef\u624b\u52a8\u4e0b\u8f7d\u3002<\/li>\n<li><strong>\u5b89\u88c5\u53ef\u9009\u4f9d\u8d56<\/strong><br \/>\n\u5982\u679c\u9700\u8981\u5fae\u8c03\u6216\u52a0\u901f\uff0c\u53ef\u5b89\u88c5\u00a0<code>torch<\/code>\u00a0\u548c\u00a0<code>datasets<\/code>\uff1a<\/li>\n<\/ol>\n<pre><code>pip install torch datasets\r\n<\/code><\/pre>\n<h3>\u5982\u4f55\u4f7f\u7528<\/h3>\n<p>Bonsai \u4f7f\u7528 Python \u811a\u672c\u8c03\u7528\u3002\u4ee5\u4e0b\u662f\u57fa\u672c\u64cd\u4f5c\u6b65\u9aa4\uff1a<\/p>\n<h4>\u52a0\u8f7d\u6a21\u578b\u548c\u5206\u8bcd\u5668<\/h4>\n<p>\u5728 Python \u4e2d\u8fd0\u884c\u4ee5\u4e0b\u4ee3\u7801\uff1a<\/p>\n<pre><code>from transformers import AutoTokenizer, AutoModelForCausalLM\r\ntokenizer = AutoTokenizer.from_pretrained(\"deepgrove\/Bonsai\", trust_remote_code=True)\r\nmodel = AutoModelForCausalLM.from_pretrained(\"deepgrove\/Bonsai\", trust_remote_code=True)\r\n<\/code><\/pre>\n<h4>\u751f\u6210\u6587\u672c<\/h4>\n<p>\u8f93\u5165\u6587\u672c\u5e76\u751f\u6210\u7ed3\u679c\uff1a<\/p>\n<pre><code>text = \"\u4e2d\u56fd\u7684\u9996\u90fd\u662f\u54ea\u91cc\uff1f\"\r\ninputs = tokenizer(text, return_tensors=\"pt\")\r\noutputs = model.generate(**inputs, max_length=100)\r\nresult = tokenizer.decode(outputs[0], skip_special_tokens=True)\r\nprint(result)\r\n<\/code><\/pre>\n<p>\u8f93\u51fa\u53ef\u80fd\u662f\u201c\u4e2d\u56fd\u7684\u9996\u90fd\u662f\u5317\u4eac\u3002\u201d\u3002<\/p>\n<h4>\u8c03\u6574\u53c2\u6570<\/h4>\n<p>\u53ef\u4fee\u6539\u751f\u6210\u53c2\u6570\uff0c\u4f8b\u5982\uff1a<\/p>\n<pre><code>outputs = model.generate(**inputs, max_length=50, temperature=0.7)\r\n<\/code><\/pre>\n<ul>\n<li><code>max_length<\/code>\uff1a\u8bbe\u7f6e\u8f93\u51fa\u957f\u5ea6\u3002<\/li>\n<li><code>temperature<\/code>\uff1a\u63a7\u5236\u8f93\u51fa\u968f\u673a\u6027\uff0c\u503c\u8d8a\u5c0f\u8d8a\u7a33\u5b9a\u3002<\/li>\n<\/ul>\n<h3>\u7279\u8272\u529f\u80fd\u64cd\u4f5c<\/h3>\n<h4>\u9ad8\u6548\u8fd0\u884c<\/h4>\n<p>Bonsai \u7684\u4e09\u503c\u6743\u91cd\u8ba9\u5b83\u5728 16 \u4f4d\u7cbe\u5ea6\u4e0b\u8fd0\u884c\u826f\u597d\u3002\u82e5\u6709 GPU\uff0c\u53ef\u81ea\u52a8\u52a0\u901f\uff1a<\/p>\n<pre><code>import torch\r\nprint(torch.cuda.is_available())  # \u8fd4\u56de True \u8868\u793a GPU \u53ef\u7528\r\n<\/code><\/pre>\n<p>GPU \u4f1a\u663e\u8457\u63d0\u5347\u6027\u80fd\uff0c\u4f46 CPU \u4e5f\u80fd\u6b63\u5e38\u8fd0\u884c\u3002<\/p>\n<h4>\u6027\u80fd\u8bc4\u4f30<\/h4>\n<p>Bonsai \u5728\u591a\u4e2a\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u8868\u73b0\u4f18\u5f02\u3002\u4ee5\u4e0b\u662f\u5b98\u65b9\u6570\u636e\uff1a<\/p>\n<table>\n<thead>\n<tr>\n<th>\u6a21\u578b<\/th>\n<th>ARC-c<\/th>\n<th>ARC-e<\/th>\n<th>HS.<\/th>\n<th>OBQA<\/th>\n<th>PiQA<\/th>\n<th>Wino.<\/th>\n<th>MMLU<\/th>\n<th>\u5e73\u5747\u5206<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>MobiLlama 0.5B<\/td>\n<td>26.62<\/td>\n<td>46.68<\/td>\n<td>51.66<\/td>\n<td>30.00<\/td>\n<td>71.65<\/td>\n<td>54.50<\/td>\n<td>28.61<\/td>\n<td>44.25<\/td>\n<\/tr>\n<tr>\n<td>Qwen 2 0.5B<\/td>\n<td>28.84<\/td>\n<td>50.29<\/td>\n<td>49.12<\/td>\n<td>33.00<\/td>\n<td>69.26<\/td>\n<td>56.99<\/td>\n<td>31.78<\/td>\n<td>45.61<\/td>\n<\/tr>\n<tr>\n<td>MobileLLM 600M<\/td>\n<td>29.01<\/td>\n<td>56.65<\/td>\n<td>55.35<\/td>\n<td>34.00<\/td>\n<td>71.65<\/td>\n<td>59.75<\/td>\n<td>31.40<\/td>\n<td>48.13<\/td>\n<\/tr>\n<tr>\n<td>Qwen 2.5 0.5B<\/td>\n<td>32.25<\/td>\n<td>58.29<\/td>\n<td>52.18<\/td>\n<td>35.40<\/td>\n<td>69.91<\/td>\n<td>56.12<\/td>\n<td>33.40<\/td>\n<td>48.22<\/td>\n<\/tr>\n<tr>\n<td><strong>Bonsai<\/strong><\/td>\n<td>33.36<\/td>\n<td>57.95<\/td>\n<td>48.04<\/td>\n<td>34.00<\/td>\n<td>70.24<\/td>\n<td>54.85<\/td>\n<td>30.28<\/td>\n<td>46.96<\/td>\n<\/tr>\n<tr>\n<td>\u8fd9\u4e9b\u6d4b\u8bd5\u5305\u62ec ARC\u3001OBQA\u3001MMLU \u7b49\uff0c\u663e\u793a Bonsai \u5728\u8f7b\u91cf\u6a21\u578b\u4e2d\u540d\u5217\u524d\u8305\u3002<\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>\u5fae\u8c03\u6a21\u578b<\/h4>\n<p>Bonsai \u672a\u7ecf\u8fc7\u6307\u4ee4\u5fae\u8c03\uff0c\u9002\u5408\u901a\u7528\u751f\u6210\u4efb\u52a1\u3002\u82e5\u9700\u4f18\u5316\u7279\u5b9a\u7528\u9014\uff08\u5982\u95ee\u7b54\uff09\uff0c\u53ef\u81ea\u884c\u5fae\u8c03\uff1a<\/p>\n<ol>\n<li>\u51c6\u5907\u6570\u636e\uff1a\u7528\u6587\u672c\u6587\u4ef6\u6216\u00a0<code>datasets<\/code>\u00a0\u5e93\u52a0\u8f7d\u3002<\/li>\n<li>\u914d\u7f6e\u53c2\u6570\uff1a\u7528\u00a0<code>TrainingArguments<\/code>\u00a0\u8bbe\u7f6e\u3002<\/li>\n<li>\u8bad\u7ec3\u6a21\u578b\uff1a<\/li>\n<\/ol>\n<pre><code>from transformers import Trainer, TrainingArguments\r\ntraining_args = TrainingArguments(\r\noutput_dir=\".\/bonsai_finetuned\",\r\nnum_train_epochs=3,\r\nper_device_train_batch_size=4\r\n)\r\ntrainer = Trainer(model=model, args=training_args, train_dataset=your_dataset)\r\ntrainer.train()\r\n<\/code><\/pre>\n<p>\u66f4\u591a\u7ec6\u8282\u89c1\u00a0<a href=\"https:\/\/huggingface.co\/docs\/transformers\/main_classes\/trainer\">Huggingface \u6587\u6863<\/a>\u3002<\/p>\n<h3>\u6ce8\u610f\u4e8b\u9879<\/h3>\n<ul>\n<li><strong>\u7cbe\u5ea6\u9650\u5236<\/strong>\uff1a\u76ee\u524d\u4ec5\u652f\u6301 16 \u4f4d\u7cbe\u5ea6\u8fd0\u884c\uff0c\u56e2\u961f\u6b63\u5f00\u53d1\u6df7\u5408\u7cbe\u5ea6\u652f\u6301\u3002<\/li>\n<li><strong>\u672a\u6307\u4ee4\u8c03\u4f18<\/strong>\uff1a\u9ed8\u8ba4\u6a21\u578b\u4e0d\u9002\u5408\u76f4\u63a5\u7528\u4e8e\u590d\u6742\u6307\u4ee4\u4efb\u52a1\uff0c\u9700\u5fae\u8c03\u3002<\/li>\n<li><strong>\u786c\u4ef6\u9700\u6c42<\/strong>\uff1a\u666e\u901a CPU \u53ef\u8fd0\u884c\uff0cGPU \u975e\u5fc5\u9700\u4f46\u63a8\u8350\u3002<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<h2>\u5e94\u7528\u573a\u666f<\/h2>\n<ol>\n<li><strong>\u6559\u80b2\u8f85\u52a9<\/strong><br \/>\nBonsai \u53ef\u56de\u7b54\u57fa\u7840\u77e5\u8bc6\u95ee\u9898\uff0c\u5982\u201c\u6cd5\u56fd\u7684\u9996\u90fd\u662f\u54ea\u91cc\uff1f\u201d\u3002\u8f93\u5165\u540e\u5feb\u901f\u751f\u6210\u7b54\u6848\uff0c\u9002\u5408\u5b66\u4e60\u4f7f\u7528\u3002<\/li>\n<li><strong>\u8fb9\u7f18\u8bbe\u5907\u5e94\u7528<\/strong><br \/>\n\u6a21\u578b\u8f7b\u91cf\uff0c\u9002\u5408\u90e8\u7f72\u5230\u624b\u673a\u6216\u5d4c\u5165\u5f0f\u8bbe\u5907\uff0c\u5b9e\u73b0\u672c\u5730\u5316\u8bed\u8a00\u5904\u7406\u3002<\/li>\n<li><strong>\u6a21\u578b\u7814\u7a76<\/strong><br \/>\n\u7814\u7a76\u8005\u53ef\u7528\u5b83\u6d4b\u8bd5\u4e09\u503c\u6743\u91cd\u6280\u672f\u7684\u6f5c\u529b\uff0c\u63a2\u7d22\u9ad8\u6548 AI \u6a21\u578b\u8bbe\u8ba1\u3002<\/li>\n<\/ol>\n<p>&nbsp;<\/p>\n<h2>QA<\/h2>\n<ol>\n<li><strong>Bonsai \u7684\u6838\u5fc3\u4f18\u52bf\u662f\u4ec0\u4e48\uff1f<\/strong><br \/>\n\u5b83\u7528\u4e09\u503c\u6743\u91cd\u6280\u672f\u5b9e\u73b0\u8f7b\u91cf\u9ad8\u6548\uff0c\u8bad\u7ec3\u6570\u636e\u5c11\u4f46\u6027\u80fd\u5f3a\uff0c\u9002\u5408\u8d44\u6e90\u53d7\u9650\u573a\u666f\u3002<\/li>\n<li><strong>\u9700\u8981 GPU \u5417\uff1f<\/strong><br \/>\n\u4e0d\u9700\u8981\u3002CPU \u5c31\u80fd\u8fd0\u884c\uff0c\u4f46 GPU \u4f1a\u52a0\u5feb\u901f\u5ea6\u3002<\/li>\n<li><strong>\u53ef\u4ee5\u76f4\u63a5\u7528\u4e8e\u5bf9\u8bdd\u5417\uff1f<\/strong><br \/>\n\u9ed8\u8ba4\u6a21\u578b\u672a\u6307\u4ee4\u8c03\u4f18\uff0c\u5efa\u8bae\u5fae\u8c03\u540e\u518d\u7528\u4e8e\u7279\u5b9a\u4efb\u52a1\u3002<\/li>\n<\/ol>\n","protected":false},"excerpt":{"rendered":"<p>Bonsai \u662f deepgrove-ai \u5f00\u53d1\u7684\u4e00\u4e2a\u5f00\u6e90\u8bed\u8a00\u6a21\u578b\uff0c\u53c2\u6570\u89c4\u6a21\u4e3a 5 \u4ebf\uff0c\u91c7\u7528\u4e09\u503c\u6743\u91cd\uff08ternary weights\uff09\u6280\u672f\u3002\u5b83\u57fa\u4e8e Llama \u67b6\u6784\u548c Mistral \u5206\u8bcd\u5668\u8bbe\u8ba1\uff0c\u7ebf\u6027\u5c42\u7ecf\u8fc7\u8c03\u6574\u4ee5\u652f\u6301\u4e09\u503c\u6743\u91cd\u3002\u6a21\u578b\u4e3b\u8981\u4f7f\u7528 &#8230;<\/p>\n","protected":false},"author":1,"featured_media":62110,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[20],"tags":[230,365],"class_list":["post-29379","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-tool","tag-aikaiyuanxiangmu","tag-damoxingweidiao"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts\/29379","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=29379"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts\/29379\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/media\/62110"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/media?parent=29379"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/categories?post=29379"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/tags?post=29379"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}