{"id":4842,"date":"2024-11-13T09:01:50","date_gmt":"2024-11-13T01:01:50","guid":{"rendered":"https:\/\/www.aqwu.net\/wp\/?p=4842"},"modified":"2024-11-13T09:53:07","modified_gmt":"2024-11-13T01:53:07","slug":"%e5%af%b9%e8%af%9d%e4%b8%ad%e5%be%ae%e8%b0%83%ef%bc%8c%e6%8f%90%e9%ab%98%e6%a8%a1%e5%9e%8b%e8%83%bd%e5%8a%9b","status":"publish","type":"post","link":"https:\/\/www.aqwu.net\/wp\/?p=4842","title":{"rendered":"\u5bf9\u8bdd\u4e2d\u5fae\u8c03\uff0c\u63d0\u9ad8\u6a21\u578b\u80fd\u529b"},"content":{"rendered":"\n<p>\u652f\u6301\u7528\u6237\u53cd\u9988\u7684\u6536\u96c6\uff0c\u5e76\u5728\u7d2f\u79ef\u8db3\u591f\u7684\u9ad8\u8d28\u91cf\u53cd\u9988\u6570\u636e\u540e\u8fdb\u884c\u5fae\u8c03\u3002\u8fd9\u4e2a\u7248\u672c\u7684\u4ee3\u7801\u4f1a\u5728\u6bcf\u6b21\u5bf9\u8bdd\u540e\uff0c\u8be2\u95ee\u7528\u6237\u662f\u5426\u6ee1\u610f\u6a21\u578b\u7684\u56de\u590d\u3002\u5982\u679c\u7528\u6237\u8868\u793a\u4e0d\u6ee1\u610f\uff0c\u8fd8\u53ef\u4ee5\u63d0\u4f9b\u66f4\u597d\u7684\u56de\u7b54\u3002\u6536\u96c6\u5230\u8db3\u591f\u7684\u53cd\u9988\u6570\u636e\u540e\uff0c\u4ee3\u7801\u4f1a\u81ea\u52a8\u89e6\u53d1\u5fae\u8c03\u3002<\/p>\n\n\n\n<p>\u4ee3\u7801\u5982\u4e0b\uff1a<\/p>\n\n\n\n<div class=\"wp-block-urvanov-syntax-highlighter-code-block\"><pre class=\"lang:python decode:true \">import json\nimport os\nfrom transformers import AutoModelForCausalLM, AutoTokenizer, Trainer, TrainingArguments\nfrom peft import get_peft_model, LoraConfig, prepare_model_for_int8_training\nimport torch\n\n# Load the model and tokenizer\nmodel_name = \"huihui-ai\/Qwen2.5-7B-Instruct-abliterated-v3\"\nmodel = AutoModelForCausalLM.from_pretrained(\n    model_name,\n    torch_dtype=\"auto\",\n    device_map=\"auto\"\n)\ntokenizer = AutoTokenizer.from_pretrained(model_name)\n\n# Configure LoRA for lightweight fine-tuning\nlora_config = LoraConfig(\n    r=8,  # Low-rank dimension\n    lora_alpha=16,\n    lora_dropout=0.1,\n    target_modules=[\"q_proj\", \"v_proj\"]  # Adjust based on model architecture\n)\nmodel = get_peft_model(model, lora_config)\n\n# Initialize or load conversation context and feedback data\ninitial_messages = [\n    {\"role\": \"system\", \"content\": \"You are Qwen, created by Alibaba Cloud. You are a helpful assistant.\"}\n]\nmessages = initial_messages.copy()\n\n# Load saved conversation and LoRA weights if they exist\nconversation_file = \"conversation_history.json\"\nlora_weights_file = \"lora_weights.pt\"\nfeedback_data_file = \"feedback_data.json\"\n\nif os.path.exists(conversation_file):\n    with open(conversation_file, \"r\") as f:\n        messages = json.load(f)\n    print(\"Loaded previous conversation history.\")\n\nif os.path.exists(lora_weights_file):\n    model.load_state_dict(torch.load(lora_weights_file), strict=False)\n    print(\"Loaded previous LoRA weights.\")\n\n# Load feedback data if it exists\nfeedback_data = []\nif os.path.exists(feedback_data_file):\n    with open(feedback_data_file, \"r\") as f:\n        feedback_data = json.load(f)\n    print(\"Loaded previous feedback data.\")\n\n# Collect user feedback and store for future fine-tuning\ndef collect_feedback(prompt, response):\n    feedback = input(\"Is this response satisfactory? (yes\/no): \").strip().lower()\n    if feedback == \"yes\":\n        feedback_data.append({\"prompt\": prompt, \"response\": response})\n    elif feedback == \"no\":\n        print(\"Feedback noted. Please provide a better response if possible.\")\n        better_response = input(\"Better response: \").strip()\n        feedback_data.append({\"prompt\": prompt, \"response\": better_response})\n\n# Fine-tune the model when enough feedback data has been collected\ndef fine_tune_with_feedback():\n    if len(feedback_data) &gt;= 10:  # Trigger fine-tuning when there are 10 feedback samples\n        print(\"Starting fine-tuning with user feedback...\")\n        \n        # Prepare dataset for training\n        train_data = [{\"input_ids\": tokenizer(feedback['prompt'], return_tensors=\"pt\").input_ids.squeeze(),\n                       \"labels\": tokenizer(feedback['response'], return_tensors=\"pt\").input_ids.squeeze()}\n                      for feedback in feedback_data]\n        \n        # Define training arguments\n        training_args = TrainingArguments(\n            output_dir=\".\/results\",\n            num_train_epochs=1,\n            per_device_train_batch_size=1,\n            save_steps=10,\n            save_total_limit=2,\n            logging_dir='.\/logs'\n        )\n        \n        trainer = Trainer(\n            model=model,\n            args=training_args,\n            train_dataset=train_data,\n        )\n        \n        trainer.train()\n        print(\"Fine-tuning complete with user feedback.\")\n        feedback_data.clear()  # Clear feedback data after fine-tuning\n\n# Enter conversation loop\ntry:\n    while True:\n        # Get user input\n        user_input = input(\"User: \").strip()\n\n        # Exit and reset commands\n        if user_input.lower() == \"\/exit\":\n            print(\"Exiting chat.\")\n            break\n        if user_input.lower() == \"\/clean\":\n            messages = initial_messages.copy()\n            print(\"Chat history cleared. Starting a new conversation.\")\n            continue\n        if not user_input:\n            print(\"Input cannot be empty. Please enter something.\")\n            continue\n\n        # Add user input to the conversation\n        messages.append({\"role\": \"user\", \"content\": user_input})\n\n        # Build the chat template\n        text = tokenizer.apply_chat_template(\n            messages,\n            tokenize=False,\n            add_generation_prompt=True\n        )\n\n        # Tokenize input and prepare it for the model\n        model_inputs = tokenizer([text], return_tensors=\"pt\").to(model.device)\n\n        # Generate a response from the model\n        generated_ids = model.generate(\n            **model_inputs,\n            max_new_tokens=150\n        )\n\n        # Extract model output, removing special tokens\n        generated_ids = [\n            output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)\n        ]\n        response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]\n\n        # Add the model's response to the conversation\n        messages.append({\"role\": \"assistant\", \"content\": response})\n\n        # Print the model's response\n        print(f\"Qwen: {response}\")\n\n        # Collect user feedback\n        collect_feedback(user_input, response)\n\n        # Fine-tune the model if there is enough feedback data\n        fine_tune_with_feedback()\n\n# Save the conversation history, LoRA weights, and feedback data when exiting\nfinally:\n    with open(conversation_file, \"w\") as f:\n        json.dump(messages, f)\n    torch.save(model.state_dict(), lora_weights_file)\n    with open(feedback_data_file, \"w\") as f:\n        json.dump(feedback_data, f)\n    print(\"Conversation history, LoRA weights, and feedback data saved.\")\n<\/pre><\/div>\n\n\n\n<h3 class=\"wp-block-heading\">\u4ee3\u7801\u89e3\u91ca<\/h3>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u53cd\u9988\u6536\u96c6<\/strong>\uff1a<code>collect_feedback<\/code> \u51fd\u6570\u4f1a\u5728\u6bcf\u6b21\u5bf9\u8bdd\u540e\u8be2\u95ee\u7528\u6237\u5bf9\u6a21\u578b\u56de\u590d\u7684\u6ee1\u610f\u5ea6\uff0c\u7528\u6237\u53ef\u4ee5\u6807\u8bb0\u201c\u6ee1\u610f\u201d\u6216\u201c\u4e0d\u6ee1\u610f\u201d\u3002\u5982\u679c\u4e0d\u6ee1\u610f\uff0c\u53ef\u4ee5\u63d0\u4f9b\u66f4\u597d\u7684\u56de\u7b54\u66ff\u4ee3\u751f\u6210\u7684\u5185\u5bb9\u3002<\/li>\n\n\n\n<li><strong>\u81ea\u52a8\u5fae\u8c03\u89e6\u53d1<\/strong>\uff1a<code>fine_tune_with_feedback<\/code> \u51fd\u6570\u4f1a\u5728\u53cd\u9988\u6570\u636e\u7d2f\u79ef\u5230\u4e00\u5b9a\u91cf\uff08\u8fd9\u91cc\u8bbe\u7f6e\u4e3a10\u6761\uff09\u65f6\u81ea\u52a8\u5fae\u8c03\u6a21\u578b\u3002\u5fae\u8c03\u540e\u7684\u6a21\u578b\u4f1a\u6839\u636e\u7528\u6237\u63d0\u4f9b\u7684\u9ad8\u8d28\u91cf\u53cd\u9988\u66f4\u597d\u5730\u54cd\u5e94\u76f8\u4f3c\u7684\u8f93\u5165\u3002<\/li>\n\n\n\n<li><strong>\u4fdd\u5b58\u548c\u52a0\u8f7d\u53cd\u9988\u6570\u636e<\/strong>\uff1a\u5728\u9000\u51fa\u65f6\uff0c\u4ee3\u7801\u4f1a\u4fdd\u5b58\u6536\u96c6\u7684\u53cd\u9988\u6570\u636e\uff0c\u4ee5\u4fbf\u4e0b\u6b21\u52a0\u8f7d\u65f6\u7ee7\u7eed\u79ef\u7d2f\uff0c\u5e76\u5728\u9700\u8981\u65f6\u518d\u6b21\u89e6\u53d1\u5fae\u8c03\u3002<\/li>\n\n\n\n<li><strong>\u4fdd\u5b58\u5fae\u8c03\u53c2\u6570<\/strong>\uff1a\u6a21\u578b\u6bcf\u6b21\u9000\u51fa\u90fd\u4f1a\u4fdd\u5b58\u6700\u65b0\u7684 LoRA \u6743\u91cd\uff0c\u8fd9\u6837\u5728\u4e0b\u6b21\u542f\u52a8\u65f6\u53ef\u4ee5\u52a0\u8f7d\u4e0a\u6b21\u7684\u5fae\u8c03\u7ed3\u679c\u3002<\/li>\n<\/ol>\n\n\n\n<h3 class=\"wp-block-heading\">\u5fae\u8c03\u540e\u9700\u8981\u91cd\u65b0\u52a0\u8f7d\u6a21\u578b\u4e48\uff1f<\/h3>\n\n\n\n<p>\u4e0d\u9700\u8981\u91cd\u65b0\u52a0\u8f7d\u6a21\u578b\uff0c\u5fae\u8c03\u8fc7\u7a0b\u76f4\u63a5\u5728\u5f53\u524d\u52a0\u8f7d\u7684\u6a21\u578b\u4e0a\u8fdb\u884c\uff0c\u66f4\u65b0\u540e\u7684\u6a21\u578b\u4f1a\u7acb\u5373\u5728\u5f53\u524d\u4f1a\u8bdd\u4e2d\u751f\u6548\u3002\u5177\u4f53\u6d41\u7a0b\u5982\u4e0b\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u5fae\u8c03\u5b8c\u6210\u540e<\/strong>\uff0c\u6a21\u578b\u53c2\u6570\u5df2\u7ecf\u66f4\u65b0\uff0c\u4e4b\u540e\u7684\u5bf9\u8bdd\u5c06\u4f1a\u4f7f\u7528\u5fae\u8c03\u540e\u7684\u6a21\u578b\u751f\u6210\u54cd\u5e94\u3002<\/li>\n\n\n\n<li><strong>\u4fdd\u5b58\u548c\u52a0\u8f7d\u6743\u91cd<\/strong>\uff1a\u9000\u51fa\u65f6\uff0c\u5c06\u5fae\u8c03\u540e\u7684\u6a21\u578b\u6743\u91cd\u4fdd\u5b58\u5230\u6587\u4ef6\u4e2d\u3002\u5982\u679c\u4f1a\u8bdd\u7ed3\u675f\u5e76\u91cd\u65b0\u542f\u52a8\uff0c\u4ee3\u7801\u4f1a\u4ece\u6587\u4ef6\u4e2d\u52a0\u8f7d\u6743\u91cd\uff0c\u4ee5\u4fdd\u6301\u4e0a\u6b21\u5fae\u8c03\u7684\u6548\u679c\u3002<\/li>\n<\/ol>\n\n\n\n<h3 class=\"wp-block-heading\">\u6ce8\u610f\u4e8b\u9879<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u9891\u7e41\u5fae\u8c03<\/strong>\uff1a\u5982\u679c\u5fae\u8c03\u8f83\u4e3a\u9891\u7e41\uff0c\u5efa\u8bae\u8bbe\u7f6e\u8f83\u4f4e\u7684\u5b66\u4e60\u7387\u6216\u9002\u5f53\u7684\u5fae\u8c03\u6b65\u6570\uff0c\u4ee5\u9632\u6a21\u578b\u8fc7\u62df\u5408\u6216\u51fa\u73b0\u4e0d\u7a33\u5b9a\u73b0\u8c61\u3002<\/li>\n\n\n\n<li><strong>\u7d2f\u79ef\u6548\u679c<\/strong>\uff1a\u5728\u957f\u65f6\u95f4\u4f7f\u7528\u540e\uff0c\u6a21\u578b\u7684\u8868\u73b0\u4f1a\u9010\u6b65\u79ef\u7d2f\u7528\u6237\u504f\u597d\u7684\u7279\u5f81\uff0c\u901a\u8fc7\u4fdd\u7559\u6743\u91cd\u6587\u4ef6\uff0c\u4fdd\u6301\u5fae\u8c03\u6548\u679c\u3002<\/li>\n<\/ul>\n\n\n\n<p><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u652f\u6301\u7528\u6237\u53cd\u9988\u7684\u6536\u96c6\uff0c\u5e76\u5728\u7d2f\u79ef\u8db3\u591f\u7684\u9ad8\u8d28\u91cf\u53cd\u9988\u6570\u636e\u540e\u8fdb\u884c\u5fae\u8c03\u3002\u8fd9\u4e2a\u7248\u672c\u7684\u4ee3\u7801\u4f1a\u5728\u6bcf\u6b21\u5bf9\u8bdd\u540e\uff0c\u8be2\u95ee\u7528\u6237\u662f\u5426\u6ee1\u610f\u6a21\u578b [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[444,443,442],"tags":[404,396],"class_list":["post-4842","post","type-post","status-publish","format-standard","hentry","category-ai","category-llm","category-llms","tag-llm","tag-396"],"views":3030,"jetpack_sharing_enabled":true,"jetpack_featured_media_url":"","_links":{"self":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts\/4842","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=4842"}],"version-history":[{"count":3,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts\/4842\/revisions"}],"predecessor-version":[{"id":4845,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts\/4842\/revisions\/4845"}],"wp:attachment":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=4842"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=4842"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=4842"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}