diff --git a/config/llm.py b/config/llm.py
new file mode 100644
index 0000000..828b9ef
--- /dev/null
+++ b/config/llm.py
@@ -0,0 +1,15 @@
+import os
+from dotenv import load_dotenv
+from typing import Dict, Any
+
+def load_config() -> Dict[str, Any]:
+ """从.env文件加载配置"""
+ load_dotenv()
+
+ return {
+ "openai": {
+ "api_key": os.getenv("OPENAI_API_KEY"),
+ "base_url": os.getenv("OPENAI_BASE_URL"),
+ "model_id": os.getenv("OPENAI_MODEL_ID")
+ }
+ }
\ No newline at end of file
diff --git a/dataset_generator.py b/dataset_generator.py
new file mode 100644
index 0000000..3f3dd2d
--- /dev/null
+++ b/dataset_generator.py
@@ -0,0 +1,94 @@
+import os
+import json
+from tools.parse_markdown import parse_markdown, MarkdownNode
+from tools.openai_api import generate_json_via_llm
+from prompt.base import create_dataset
+from config.llm import load_config
+from tqdm import tqdm
+
+def process_markdown_file(file_path):
+ with open(file_path, 'r', encoding='utf-8') as f:
+ content = f.read()
+
+ root = parse_markdown(content)
+ results = []
+
+ def traverse(node, parent_titles):
+ current_titles = parent_titles.copy()
+ current_titles.append(node.title)
+
+ if not node.children: # 叶子节点
+ if node.content:
+ full_text = ' -> '.join(current_titles) + '\n' + node.content
+ results.append(full_text)
+ else:
+ for child in node.children:
+ traverse(child, current_titles)
+
+ traverse(root, [])
+ return results
+
+def find_markdown_files(directory):
+ markdown_files = []
+ for root, dirs, files in os.walk(directory):
+ for file in files:
+ if file.endswith('.md'):
+ markdown_files.append(os.path.join(root, file))
+ return markdown_files
+
+def process_all_markdown(doc_dir):
+ all_results = []
+
+ markdown_files = find_markdown_files(doc_dir)
+ for file_path in markdown_files:
+ results = process_markdown_file(file_path)
+ all_results.extend(results)
+
+ return all_results
+
+def save_dataset(dataset, output_dir):
+ os.makedirs(output_dir, exist_ok=True)
+ output_path = os.path.join(output_dir, 'dataset.json')
+ with open(output_path, 'w', encoding='utf-8') as f:
+ json.dump(dataset, f, ensure_ascii=False, indent=2)
+
+if __name__ == "__main__":
+ # 解析markdown文档
+ results = process_all_markdown('workdir/my_docs')
+
+ # 加载LLM配置
+ config = load_config()
+
+ dataset = []
+ # 使用tqdm包装外部循环以显示进度条
+ for content in tqdm(results, desc="生成数据集进度", unit="文档"):
+ for _ in range(3):
+ prompt = create_dataset.create(
+ "LLaMA-Factory", # 项目名
+ content, # 文档内容
+ """{
+ "dataset":[
+ {
+ "question":"",
+ "answer":""
+ }
+ ]
+ }"""
+ )
+
+ # 调用LLM生成JSON
+ try:
+ result = generate_json_via_llm(
+ prompt=prompt,
+ base_url=config["openai"]["base_url"],
+ api_key=config["openai"]["api_key"],
+ model_id=config["openai"]["model_id"]
+ )
+ print(json.loads(result)["dataset"])
+ dataset.extend(json.loads(result)["dataset"])
+ except Exception as e:
+ print(f"生成数据集时出错: {e}")
+
+ # 保存数据集
+ save_dataset(dataset, 'workdir/dataset2')
+ print(f"数据集已生成，共{len(dataset)}条数据")
\ No newline at end of file
diff --git a/prompt/base.py b/prompt/base.py
new file mode 100644
index 0000000..9c06d9e
--- /dev/null
+++ b/prompt/base.py
@@ -0,0 +1,25 @@
+class create_dataset:
+ """用于生成微调数据集模板的类"""
+
+ template = """
+项目名为:{}
+请依据以下该项目官方文档的部分内容，创造合适的对话数据集用于微调一个了解该项目的小模型的语料，要求兼顾文档中间尽可能多的信息点，使用中文
+文档节选：{}
+按照如下json格式返回：{}
+"""
+
+ @staticmethod
+ def create(*args: any) -> str:
+ """根据提供的任意数量参数生成数据集模板
+
+ Args:
+ *args: 任意数量的参数，将按顺序填充到模板中
+
+ Returns:
+ 格式化后的模板字符串
+ """
+ return create_dataset.template.format(*args)
+
+
+if __name__=="__main__":
+ print(create_dataset.create("a", "b", "c"))
\ No newline at end of file
diff --git a/schema/dataset.py b/schema/dataset.py
new file mode 100644
index 0000000..5d27d95
--- /dev/null
+++ b/schema/dataset.py
@@ -0,0 +1,9 @@
+from pydantic import BaseModel, RootModel
+from typing import List
+
+class QAPair(BaseModel):
+ question: str
+ response: str
+
+class QAArray(RootModel):
+ root: List[QAPair]
\ No newline at end of file
diff --git a/tools/openai_api.py b/tools/openai_api.py
new file mode 100644
index 0000000..5f21462
--- /dev/null
+++ b/tools/openai_api.py
@@ -0,0 +1,69 @@
+import json
+from openai import OpenAI
+
+def generate_json_via_llm(
+ prompt: str,
+ base_url: str,
+ api_key: str,
+ model_id: str
+) -> str:
+ client = OpenAI(
+ api_key=api_key,
+ base_url=base_url
+ )
+
+ try:
+
+
+ response = client.chat.completions.create(
+ model=model_id,
+ messages=[
+ {
+ "role": "user",
+ "content": prompt
+ }
+ ],
+ response_format={
+ 'type': 'json_object'
+ }
+ )
+
+ return response.choices[0].message.content
+ except Exception as e:
+ raise RuntimeError(f"API请求失败: {e}")
+
+
+if __name__ == "__main__":
+ import sys
+ import os
+ sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), '..')))
+ from config.llm import load_config
+ # 将项目根目录添加到 sys.path 中
+
+ # 示例用法
+ try:
+ config = load_config()
+ print(config)
+ result = generate_json_via_llm(
+ prompt="""测试，随便生成点什么，返回json格式的字符串,格式如下
+{
+ "dataset":[
+ {
+ "question":"",
+ "answer":""
+ },
+ {
+ "question":"",
+ "answer":""
+ }
+ ......
+ ]
+}
+ """,
+ base_url=config["openai"]["base_url"],
+ api_key=config["openai"]["api_key"],
+ model_id=config["openai"]["model_id"],
+ )
+ print(result)
+ except Exception as e:
+ print(f"错误: {e}")
diff --git a/train.ipynb b/train.ipynb
new file mode 100644
index 0000000..c539f87
--- /dev/null
+++ b/train.ipynb
@@ -0,0 +1,1534 @@
+{
+ "cells": [
+ {
+ "cell_type": "code",
+ "execution_count": 1,
+ "metadata": {},
+ "outputs": [
+ {
+ "name": "stdout",
+ "output_type": "stream",
+ "text": [
+ "🦥 Unsloth: Will patch your computer to enable 2x faster free finetuning.\n"
+ ]
+ },
+ {
+ "name": "stderr",
+ "output_type": "stream",
+ "text": [
+ "c:\\Users\\sbtwc\\.conda\\envs\\unsloth_env\\Lib\\site-packages\\tqdm\\auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html\n",
+ " from .autonotebook import tqdm as notebook_tqdm\n"
+ ]
+ },
+ {
+ "name": "stdout",
+ "output_type": "stream",
+ "text": [
+ "Unsloth: OpenAI failed to import - ignoring for now.\n",
+ "🦥 Unsloth Zoo will now patch everything to make training faster!\n"
+ ]
+ },
+ {
+ "name": "stderr",
+ "output_type": "stream",
+ "text": [
+ "c:\\Users\\sbtwc\\.conda\\envs\\unsloth_env\\Lib\\site-packages\\unsloth_zoo\\gradient_checkpointing.py:330: UserWarning: expandable_segments not supported on this platform (Triggered internally at C:\\actions-runner\\_work\\pytorch\\pytorch\\pytorch\\c10/cuda/CUDAAllocatorConfig.h:28.)\n",
+ " GPU_BUFFERS = tuple([torch.empty(2*256*2048, dtype = dtype, device = f\"cuda:{i}\") for i in range(n_gpus)])\n"
+ ]
+ },
+ {
+ "name": "stdout",
+ "output_type": "stream",
+ "text": [
+ "==((====))== Unsloth 2025.3.9: Fast Qwen2 patching. Transformers: 4.48.1.\n",
+ " \\\\ /| NVIDIA GeForce RTX 3060 Laptop GPU. Num GPUs = 1. Max memory: 6.0 GB. Platform: Windows.\n",
+ "O^O/ \\_/ \\ Torch: 2.6.0+cu126. CUDA: 8.6. CUDA Toolkit: 12.6. Triton: 3.2.0\n",
+ "\\ / Bfloat16 = TRUE. FA [Xformers = 0.0.29.post3. FA2 = False]\n",
+ " \"-____-\" Free license: http://github.com/unslothai/unsloth\n",
+ "Unsloth: Fast downloading is enabled - ignore downloading bars which are red colored!\n"
+ ]
+ },
+ {
+ "name": "stderr",
+ "output_type": "stream",
+ "text": [
+ "Unsloth 2025.3.9 patched 36 layers with 36 QKV layers, 36 O layers and 36 MLP layers.\n"
+ ]
+ }
+ ],
+ "source": [
+ "from unsloth import FastLanguageModel\n",
+ "import torch\n",
+ "\n",
+ "# 基础配置参数\n",
+ "max_seq_length = 4096 # 最大序列长度\n",
+ "dtype = None # 自动检测数据类型\n",
+ "load_in_4bit = True # 使用4位量化以减少内存使用\n",
+ "\n",
+ "# 加载预训练模型和分词器\n",
+ "model, tokenizer = FastLanguageModel.from_pretrained(\n",
+ " model_name = \"workdir/model/Qwen2.5-3B-Instruct-bnb-4bit\", # 选择Qwen2.5 32B指令模型\n",
+ " max_seq_length = max_seq_length,\n",
+ " dtype = dtype,\n",
+ " load_in_4bit = load_in_4bit,\n",
+ ")\n",
+ "\n",
+ "model = FastLanguageModel.get_peft_model(\n",
+ " model,\n",
+ " r = 64, # LoRA秩,控制可训练参数数量\n",
+ " target_modules = [\"q_proj\", \"k_proj\", \"v_proj\", \"o_proj\",\n",
+ " \"gate_proj\", \"up_proj\", \"down_proj\",], # 需要训练的目标模块\n",
+ " lora_alpha = 64, # LoRA缩放因子\n",
+ " lora_dropout = 0, # LoRA dropout率\n",
+ " bias = \"none\", # 是否训练偏置项\n",
+ " use_gradient_checkpointing = \"unsloth\", # 使用梯度检查点节省显存\n",
+ " random_state = 114514, # 随机数种子\n",
+ " use_rslora = False, # 是否使用稳定版LoRA\n",
+ " loftq_config = None, # LoftQ配置\n",
+ ")\n",
+ "\n",
+ "from unsloth.chat_templates import get_chat_template\n",
+ "# 配置分词器使用qwen-2.5对话模板\n",
+ "tokenizer = get_chat_template(\n",
+ " tokenizer,\n",
+ " chat_template=\"qwen-2.5\",\n",
+ ")\n"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "metadata": {},
+ "outputs": [
+ {
+ "name": "stderr",
+ "output_type": "stream",
+ "text": [
+ "Map: 100%|██████████| 595/595 [00:00<00:00, 7000.21 examples/s]\n"
+ ]
+ }
+ ],
+ "source": [
+ "# 加载数据集\n",
+ "import json\n",
+ "\n",
+ "def formatting_prompts_func(examples):\n",
+ " \"\"\"格式化对话数据的函数\n",
+ " Args:\n",
+ " examples: 包含对话列表的字典\n",
+ " Returns:\n",
+ " 包含格式化文本的字典\n",
+ " \"\"\"\n",
+ " questions = examples[\"question\"]\n",
+ " answer = examples[\"answer\"]\n",
+ " \n",
+ " # 将Question和Response组合成对话形式\n",
+ " convos = [\n",
+ " [{\"role\": \"user\", \"content\": q}, {\"role\": \"assistant\", \"content\": r}]\n",
+ " for q, r in zip(questions, answer)\n",
+ " ]\n",
+ " \n",
+ " # 使用tokenizer.apply_chat_template格式化对话\n",
+ " texts = [\n",
+ " tokenizer.apply_chat_template(convo, tokenize=False, add_generation_prompt=False)\n",
+ " for convo in convos\n",
+ " ]\n",
+ " \n",
+ " return {\"text\": texts}\n",
+ "\n",
+ "\n"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 10,
+ "metadata": {},
+ "outputs": [
+ {
+ "name": "stdout",
+ "output_type": "stream",
+ "text": [
+ "{'question': 'LLaMA-Factory有哪些训练方法？', 'answer': 'LLaMA-Factory提供了Pre-training和Post-training两种训练方法。', 'text': '<|im_start|>system\\nYou are Qwen, created by Alibaba Cloud. You are a helpful assistant.<|im_end|>\\n<|im_start|>user\\nLLaMA-Factory有哪些训练方法？<|im_end|>\\n<|im_start|>assistant\\nLLaMA-Factory提供了Pre-training和Post-training两种训练方法。<|im_end|>\\n'}\n"
+ ]
+ }
+ ],
+ "source": [
+ "from datasets import load_dataset\n",
+ "dataset = load_dataset(\"json\", data_files=\"workdir\\dataset\\dataset.json\",split=\"train\")\n",
+ "dataset = dataset.map(formatting_prompts_func, batched = True)\n",
+ "\n",
+ "print(dataset[5])"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 13,
+ "metadata": {},
+ "outputs": [
+ {
+ "name": "stdout",
+ "output_type": "stream",
+ "text": [
+ "GPU = NVIDIA GeForce RTX 3060 Laptop GPU. Max memory = 6.0 GB.\n",
+ "2.557 GB of memory reserved.\n"
+ ]
+ }
+ ],
+ "source": [
+ "from trl import SFTTrainer\n",
+ "from transformers import TrainingArguments, DataCollatorForSeq2Seq\n",
+ "from unsloth import is_bfloat16_supported\n",
+ "\n",
+ "# 配置训练器\n",
+ "trainer = SFTTrainer(\n",
+ " model=model,\n",
+ " tokenizer=tokenizer,\n",
+ " train_dataset=dataset,\n",
+ " dataset_text_field=\"text\",\n",
+ " max_seq_length=max_seq_length,\n",
+ " data_collator=DataCollatorForSeq2Seq(tokenizer=tokenizer),\n",
+ " dataset_num_proc=1,\n",
+ " packing=False,\n",
+ " args=TrainingArguments(\n",
+ " per_device_train_batch_size=1, # 每个设备的批次大小\n",
+ " gradient_accumulation_steps=4, # 梯度累积步数\n",
+ " warmup_steps=3*50, # 预热步数\n",
+ " max_steps=3*500, # 最大训练步数\n",
+ " learning_rate=1e-4, # 学习率\n",
+ " fp16=not is_bfloat16_supported(), # 是否使用fp16\n",
+ " bf16=is_bfloat16_supported(), # 是否使用bf16\n",
+ " logging_steps=5, # 日志记录间隔\n",
+ " optim=\"paged_adamw_8bit\", # 优化器\n",
+ " weight_decay=0.01, # 权重衰减\n",
+ " lr_scheduler_type=\"linear\", # 学习率调度器\n",
+ " seed=114514, # 随机种子\n",
+ " output_dir=\"workdir/checkpoint/\", # 输出目录\n",
+ " save_strategy=\"steps\", # 按步保存中间权重\n",
+ " save_steps=200, # 每多少步保存一次中间权重\n",
+ " report_to=\"none\", # 不使用外部日志工具\n",
+ " ),\n",
+ ")\n",
+ "\n",
+ "from unsloth.chat_templates import train_on_responses_only\n",
+ "# 设置仅对助手回复部分计算损失\n",
+ "trainer = train_on_responses_only(\n",
+ " trainer,\n",
+ " instruction_part = \"<|im_start|>user\\n\",\n",
+ " response_part = \"<|im_start|>assistant\\n\",\n",
+ ")\n",
+ "\n",
+ "# 获取GPU信息\n",
+ "gpu_stats = torch.cuda.get_device_properties(0)\n",
+ "start_gpu_memory = round(torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024, 3)\n",
+ "max_memory = round(gpu_stats.total_memory / 1024 / 1024 / 1024, 3)\n",
+ "print(f\"GPU = {gpu_stats.name}. Max memory = {max_memory} GB.\")\n",
+ "print(f\"{start_gpu_memory} GB of memory reserved.\")"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 15,
+ "metadata": {},
+ "outputs": [
+ {
+ "name": "stderr",
+ "output_type": "stream",
+ "text": [
+ "==((====))== Unsloth - 2x faster free finetuning | Num GPUs used = 1\n",
+ " \\\\ /| Num examples = 595 | Num Epochs = 11 | Total steps = 1,500\n",
+ "O^O/ \\_/ \\ Batch size per device = 1 | Gradient accumulation steps = 4\n",
+ "\\ / Data Parallel GPUs = 1 | Total batch size (1 x 4 x 1) = 4\n",
+ " \"-____-\" Trainable parameters = 119,734,272/1,818,406,912 (6.58% trained)\n"
+ ]
+ },
+ {
+ "data": {
+ "text/html": [
+ "\n",
+ " \n",
+ " \n",
+ "
\n",
+ " [1500/1500 1:15:43, Epoch 10/11]\n",
+ "
\n",
+ " \n",
+ " \n",
+ " | Step | \n",
+ " Training Loss | \n",
+ "
\n",
+ " \n",
+ " \n",
+ " \n",
+ " | 5 | \n",
+ " 2.941600 | \n",
+ "
\n",
+ " \n",
+ " | 10 | \n",
+ " 2.629000 | \n",
+ "
\n",
+ " \n",
+ " | 15 | \n",
+ " 2.573900 | \n",
+ "
\n",
+ " \n",
+ " | 20 | \n",
+ " 1.995600 | \n",
+ "
\n",
+ " \n",
+ " | 25 | \n",
+ " 1.651400 | \n",
+ "
\n",
+ " \n",
+ " | 30 | \n",
+ " 1.505400 | \n",
+ "
\n",
+ " \n",
+ " | 35 | \n",
+ " 1.709300 | \n",
+ "
\n",
+ " \n",
+ " | 40 | \n",
+ " 1.530200 | \n",
+ "
\n",
+ " \n",
+ " | 45 | \n",
+ " 1.362000 | \n",
+ "
\n",
+ " \n",
+ " | 50 | \n",
+ " 1.413000 | \n",
+ "
\n",
+ " \n",
+ " | 55 | \n",
+ " 1.291000 | \n",
+ "
\n",
+ " \n",
+ " | 60 | \n",
+ " 1.365500 | \n",
+ "
\n",
+ " \n",
+ " | 65 | \n",
+ " 1.374200 | \n",
+ "
\n",
+ " \n",
+ " | 70 | \n",
+ " 1.313900 | \n",
+ "
\n",
+ " \n",
+ " | 75 | \n",
+ " 1.388400 | \n",
+ "
\n",
+ " \n",
+ " | 80 | \n",
+ " 1.292300 | \n",
+ "
\n",
+ " \n",
+ " | 85 | \n",
+ " 1.205700 | \n",
+ "
\n",
+ " \n",
+ " | 90 | \n",
+ " 1.162200 | \n",
+ "
\n",
+ " \n",
+ " | 95 | \n",
+ " 1.194100 | \n",
+ "
\n",
+ " \n",
+ " | 100 | \n",
+ " 0.905200 | \n",
+ "
\n",
+ " \n",
+ " | 105 | \n",
+ " 1.107500 | \n",
+ "
\n",
+ " \n",
+ " | 110 | \n",
+ " 0.915300 | \n",
+ "
\n",
+ " \n",
+ " | 115 | \n",
+ " 1.197300 | \n",
+ "
\n",
+ " \n",
+ " | 120 | \n",
+ " 0.832900 | \n",
+ "
\n",
+ " \n",
+ " | 125 | \n",
+ " 1.005700 | \n",
+ "
\n",
+ " \n",
+ " | 130 | \n",
+ " 0.883100 | \n",
+ "
\n",
+ " \n",
+ " | 135 | \n",
+ " 1.002800 | \n",
+ "
\n",
+ " \n",
+ " | 140 | \n",
+ " 0.871200 | \n",
+ "
\n",
+ " \n",
+ " | 145 | \n",
+ " 0.896200 | \n",
+ "
\n",
+ " \n",
+ " | 150 | \n",
+ " 0.738400 | \n",
+ "
\n",
+ " \n",
+ " | 155 | \n",
+ " 0.821300 | \n",
+ "
\n",
+ " \n",
+ " | 160 | \n",
+ " 0.540100 | \n",
+ "
\n",
+ " \n",
+ " | 165 | \n",
+ " 0.528400 | \n",
+ "
\n",
+ " \n",
+ " | 170 | \n",
+ " 0.748300 | \n",
+ "
\n",
+ " \n",
+ " | 175 | \n",
+ " 0.530400 | \n",
+ "
\n",
+ " \n",
+ " | 180 | \n",
+ " 0.827400 | \n",
+ "
\n",
+ " \n",
+ " | 185 | \n",
+ " 0.462000 | \n",
+ "
\n",
+ " \n",
+ " | 190 | \n",
+ " 0.662100 | \n",
+ "
\n",
+ " \n",
+ " | 195 | \n",
+ " 0.460600 | \n",
+ "
\n",
+ " \n",
+ " | 200 | \n",
+ " 0.489000 | \n",
+ "
\n",
+ " \n",
+ " | 205 | \n",
+ " 0.718200 | \n",
+ "
\n",
+ " \n",
+ " | 210 | \n",
+ " 0.560800 | \n",
+ "
\n",
+ " \n",
+ " | 215 | \n",
+ " 0.465400 | \n",
+ "
\n",
+ " \n",
+ " | 220 | \n",
+ " 0.563800 | \n",
+ "
\n",
+ " \n",
+ " | 225 | \n",
+ " 0.511300 | \n",
+ "
\n",
+ " \n",
+ " | 230 | \n",
+ " 0.633600 | \n",
+ "
\n",
+ " \n",
+ " | 235 | \n",
+ " 0.672000 | \n",
+ "
\n",
+ " \n",
+ " | 240 | \n",
+ " 0.512300 | \n",
+ "
\n",
+ " \n",
+ " | 245 | \n",
+ " 0.435900 | \n",
+ "
\n",
+ " \n",
+ " | 250 | \n",
+ " 0.602000 | \n",
+ "
\n",
+ " \n",
+ " | 255 | \n",
+ " 0.410400 | \n",
+ "
\n",
+ " \n",
+ " | 260 | \n",
+ " 0.444500 | \n",
+ "
\n",
+ " \n",
+ " | 265 | \n",
+ " 0.498200 | \n",
+ "
\n",
+ " \n",
+ " | 270 | \n",
+ " 0.474100 | \n",
+ "
\n",
+ " \n",
+ " | 275 | \n",
+ " 0.499900 | \n",
+ "
\n",
+ " \n",
+ " | 280 | \n",
+ " 0.472900 | \n",
+ "
\n",
+ " \n",
+ " | 285 | \n",
+ " 0.515300 | \n",
+ "
\n",
+ " \n",
+ " | 290 | \n",
+ " 0.710300 | \n",
+ "
\n",
+ " \n",
+ " | 295 | \n",
+ " 0.471400 | \n",
+ "
\n",
+ " \n",
+ " | 300 | \n",
+ " 0.435600 | \n",
+ "
\n",
+ " \n",
+ " | 305 | \n",
+ " 0.227500 | \n",
+ "
\n",
+ " \n",
+ " | 310 | \n",
+ " 0.329300 | \n",
+ "
\n",
+ " \n",
+ " | 315 | \n",
+ " 0.278500 | \n",
+ "
\n",
+ " \n",
+ " | 320 | \n",
+ " 0.199700 | \n",
+ "
\n",
+ " \n",
+ " | 325 | \n",
+ " 0.185300 | \n",
+ "
\n",
+ " \n",
+ " | 330 | \n",
+ " 0.319400 | \n",
+ "
\n",
+ " \n",
+ " | 335 | \n",
+ " 0.221700 | \n",
+ "
\n",
+ " \n",
+ " | 340 | \n",
+ " 0.199100 | \n",
+ "
\n",
+ " \n",
+ " | 345 | \n",
+ " 0.263700 | \n",
+ "
\n",
+ " \n",
+ " | 350 | \n",
+ " 0.162200 | \n",
+ "
\n",
+ " \n",
+ " | 355 | \n",
+ " 0.240800 | \n",
+ "
\n",
+ " \n",
+ " | 360 | \n",
+ " 0.233900 | \n",
+ "
\n",
+ " \n",
+ " | 365 | \n",
+ " 0.283800 | \n",
+ "
\n",
+ " \n",
+ " | 370 | \n",
+ " 0.234300 | \n",
+ "
\n",
+ " \n",
+ " | 375 | \n",
+ " 0.280000 | \n",
+ "
\n",
+ " \n",
+ " | 380 | \n",
+ " 0.421600 | \n",
+ "
\n",
+ " \n",
+ " | 385 | \n",
+ " 0.244700 | \n",
+ "
\n",
+ " \n",
+ " | 390 | \n",
+ " 0.263500 | \n",
+ "
\n",
+ " \n",
+ " | 395 | \n",
+ " 0.227000 | \n",
+ "
\n",
+ " \n",
+ " | 400 | \n",
+ " 0.200800 | \n",
+ "
\n",
+ " \n",
+ " | 405 | \n",
+ " 0.196800 | \n",
+ "
\n",
+ " \n",
+ " | 410 | \n",
+ " 0.226100 | \n",
+ "
\n",
+ " \n",
+ " | 415 | \n",
+ " 0.267700 | \n",
+ "
\n",
+ " \n",
+ " | 420 | \n",
+ " 0.166400 | \n",
+ "
\n",
+ " \n",
+ " | 425 | \n",
+ " 0.307700 | \n",
+ "
\n",
+ " \n",
+ " | 430 | \n",
+ " 0.295600 | \n",
+ "
\n",
+ " \n",
+ " | 435 | \n",
+ " 0.184200 | \n",
+ "
\n",
+ " \n",
+ " | 440 | \n",
+ " 0.196100 | \n",
+ "
\n",
+ " \n",
+ " | 445 | \n",
+ " 0.220000 | \n",
+ "
\n",
+ " \n",
+ " | 450 | \n",
+ " 0.137200 | \n",
+ "
\n",
+ " \n",
+ " | 455 | \n",
+ " 0.115100 | \n",
+ "
\n",
+ " \n",
+ " | 460 | \n",
+ " 0.157100 | \n",
+ "
\n",
+ " \n",
+ " | 465 | \n",
+ " 0.142500 | \n",
+ "
\n",
+ " \n",
+ " | 470 | \n",
+ " 0.161300 | \n",
+ "
\n",
+ " \n",
+ " | 475 | \n",
+ " 0.121700 | \n",
+ "
\n",
+ " \n",
+ " | 480 | \n",
+ " 0.182200 | \n",
+ "
\n",
+ " \n",
+ " | 485 | \n",
+ " 0.094400 | \n",
+ "
\n",
+ " \n",
+ " | 490 | \n",
+ " 0.135800 | \n",
+ "
\n",
+ " \n",
+ " | 495 | \n",
+ " 0.115100 | \n",
+ "
\n",
+ " \n",
+ " | 500 | \n",
+ " 0.144500 | \n",
+ "
\n",
+ " \n",
+ " | 505 | \n",
+ " 0.148100 | \n",
+ "
\n",
+ " \n",
+ " | 510 | \n",
+ " 0.099800 | \n",
+ "
\n",
+ " \n",
+ " | 515 | \n",
+ " 0.131800 | \n",
+ "
\n",
+ " \n",
+ " | 520 | \n",
+ " 0.150100 | \n",
+ "
\n",
+ " \n",
+ " | 525 | \n",
+ " 0.130300 | \n",
+ "
\n",
+ " \n",
+ " | 530 | \n",
+ " 0.153200 | \n",
+ "
\n",
+ " \n",
+ " | 535 | \n",
+ " 0.178000 | \n",
+ "
\n",
+ " \n",
+ " | 540 | \n",
+ " 0.107700 | \n",
+ "
\n",
+ " \n",
+ " | 545 | \n",
+ " 0.182500 | \n",
+ "
\n",
+ " \n",
+ " | 550 | \n",
+ " 0.151500 | \n",
+ "
\n",
+ " \n",
+ " | 555 | \n",
+ " 0.157400 | \n",
+ "
\n",
+ " \n",
+ " | 560 | \n",
+ " 0.186700 | \n",
+ "
\n",
+ " \n",
+ " | 565 | \n",
+ " 0.192900 | \n",
+ "
\n",
+ " \n",
+ " | 570 | \n",
+ " 0.137400 | \n",
+ "
\n",
+ " \n",
+ " | 575 | \n",
+ " 0.099100 | \n",
+ "
\n",
+ " \n",
+ " | 580 | \n",
+ " 0.094500 | \n",
+ "
\n",
+ " \n",
+ " | 585 | \n",
+ " 0.117100 | \n",
+ "
\n",
+ " \n",
+ " | 590 | \n",
+ " 0.150600 | \n",
+ "
\n",
+ " \n",
+ " | 595 | \n",
+ " 0.093600 | \n",
+ "
\n",
+ " \n",
+ " | 600 | \n",
+ " 0.090400 | \n",
+ "
\n",
+ " \n",
+ " | 605 | \n",
+ " 0.068400 | \n",
+ "
\n",
+ " \n",
+ " | 610 | \n",
+ " 0.107400 | \n",
+ "
\n",
+ " \n",
+ " | 615 | \n",
+ " 0.034200 | \n",
+ "
\n",
+ " \n",
+ " | 620 | \n",
+ " 0.075000 | \n",
+ "
\n",
+ " \n",
+ " | 625 | \n",
+ " 0.073900 | \n",
+ "
\n",
+ " \n",
+ " | 630 | \n",
+ " 0.078400 | \n",
+ "
\n",
+ " \n",
+ " | 635 | \n",
+ " 0.077900 | \n",
+ "
\n",
+ " \n",
+ " | 640 | \n",
+ " 0.065600 | \n",
+ "
\n",
+ " \n",
+ " | 645 | \n",
+ " 0.101700 | \n",
+ "
\n",
+ " \n",
+ " | 650 | \n",
+ " 0.084900 | \n",
+ "
\n",
+ " \n",
+ " | 655 | \n",
+ " 0.073000 | \n",
+ "
\n",
+ " \n",
+ " | 660 | \n",
+ " 0.100800 | \n",
+ "
\n",
+ " \n",
+ " | 665 | \n",
+ " 0.035700 | \n",
+ "
\n",
+ " \n",
+ " | 670 | \n",
+ " 0.076300 | \n",
+ "
\n",
+ " \n",
+ " | 675 | \n",
+ " 0.077500 | \n",
+ "
\n",
+ " \n",
+ " | 680 | \n",
+ " 0.060200 | \n",
+ "
\n",
+ " \n",
+ " | 685 | \n",
+ " 0.107900 | \n",
+ "
\n",
+ " \n",
+ " | 690 | \n",
+ " 0.109300 | \n",
+ "
\n",
+ " \n",
+ " | 695 | \n",
+ " 0.082700 | \n",
+ "
\n",
+ " \n",
+ " | 700 | \n",
+ " 0.075900 | \n",
+ "
\n",
+ " \n",
+ " | 705 | \n",
+ " 0.088300 | \n",
+ "
\n",
+ " \n",
+ " | 710 | \n",
+ " 0.112000 | \n",
+ "
\n",
+ " \n",
+ " | 715 | \n",
+ " 0.084100 | \n",
+ "
\n",
+ " \n",
+ " | 720 | \n",
+ " 0.127700 | \n",
+ "
\n",
+ " \n",
+ " | 725 | \n",
+ " 0.070700 | \n",
+ "
\n",
+ " \n",
+ " | 730 | \n",
+ " 0.085400 | \n",
+ "
\n",
+ " \n",
+ " | 735 | \n",
+ " 0.054400 | \n",
+ "
\n",
+ " \n",
+ " | 740 | \n",
+ " 0.083300 | \n",
+ "
\n",
+ " \n",
+ " | 745 | \n",
+ " 0.044600 | \n",
+ "
\n",
+ " \n",
+ " | 750 | \n",
+ " 0.025700 | \n",
+ "
\n",
+ " \n",
+ " | 755 | \n",
+ " 0.039400 | \n",
+ "
\n",
+ " \n",
+ " | 760 | \n",
+ " 0.056600 | \n",
+ "
\n",
+ " \n",
+ " | 765 | \n",
+ " 0.050800 | \n",
+ "
\n",
+ " \n",
+ " | 770 | \n",
+ " 0.042500 | \n",
+ "
\n",
+ " \n",
+ " | 775 | \n",
+ " 0.054000 | \n",
+ "
\n",
+ " \n",
+ " | 780 | \n",
+ " 0.061200 | \n",
+ "
\n",
+ " \n",
+ " | 785 | \n",
+ " 0.064100 | \n",
+ "
\n",
+ " \n",
+ " | 790 | \n",
+ " 0.048600 | \n",
+ "
\n",
+ " \n",
+ " | 795 | \n",
+ " 0.048600 | \n",
+ "
\n",
+ " \n",
+ " | 800 | \n",
+ " 0.050300 | \n",
+ "
\n",
+ " \n",
+ " | 805 | \n",
+ " 0.056400 | \n",
+ "
\n",
+ " \n",
+ " | 810 | \n",
+ " 0.051000 | \n",
+ "
\n",
+ " \n",
+ " | 815 | \n",
+ " 0.060900 | \n",
+ "
\n",
+ " \n",
+ " | 820 | \n",
+ " 0.054600 | \n",
+ "
\n",
+ " \n",
+ " | 825 | \n",
+ " 0.024800 | \n",
+ "
\n",
+ " \n",
+ " | 830 | \n",
+ " 0.027800 | \n",
+ "
\n",
+ " \n",
+ " | 835 | \n",
+ " 0.083900 | \n",
+ "
\n",
+ " \n",
+ " | 840 | \n",
+ " 0.046700 | \n",
+ "
\n",
+ " \n",
+ " | 845 | \n",
+ " 0.073400 | \n",
+ "
\n",
+ " \n",
+ " | 850 | \n",
+ " 0.030800 | \n",
+ "
\n",
+ " \n",
+ " | 855 | \n",
+ " 0.059400 | \n",
+ "
\n",
+ " \n",
+ " | 860 | \n",
+ " 0.027300 | \n",
+ "
\n",
+ " \n",
+ " | 865 | \n",
+ " 0.066000 | \n",
+ "
\n",
+ " \n",
+ " | 870 | \n",
+ " 0.080000 | \n",
+ "
\n",
+ " \n",
+ " | 875 | \n",
+ " 0.059600 | \n",
+ "
\n",
+ " \n",
+ " | 880 | \n",
+ " 0.052600 | \n",
+ "
\n",
+ " \n",
+ " | 885 | \n",
+ " 0.055900 | \n",
+ "
\n",
+ " \n",
+ " | 890 | \n",
+ " 0.042300 | \n",
+ "
\n",
+ " \n",
+ " | 895 | \n",
+ " 0.034500 | \n",
+ "
\n",
+ " \n",
+ " | 900 | \n",
+ " 0.019600 | \n",
+ "
\n",
+ " \n",
+ " | 905 | \n",
+ " 0.027800 | \n",
+ "
\n",
+ " \n",
+ " | 910 | \n",
+ " 0.012100 | \n",
+ "
\n",
+ " \n",
+ " | 915 | \n",
+ " 0.027300 | \n",
+ "
\n",
+ " \n",
+ " | 920 | \n",
+ " 0.036900 | \n",
+ "
\n",
+ " \n",
+ " | 925 | \n",
+ " 0.030100 | \n",
+ "
\n",
+ " \n",
+ " | 930 | \n",
+ " 0.027900 | \n",
+ "
\n",
+ " \n",
+ " | 935 | \n",
+ " 0.028700 | \n",
+ "
\n",
+ " \n",
+ " | 940 | \n",
+ " 0.061500 | \n",
+ "
\n",
+ " \n",
+ " | 945 | \n",
+ " 0.025500 | \n",
+ "
\n",
+ " \n",
+ " | 950 | \n",
+ " 0.020100 | \n",
+ "
\n",
+ " \n",
+ " | 955 | \n",
+ " 0.021700 | \n",
+ "
\n",
+ " \n",
+ " | 960 | \n",
+ " 0.026800 | \n",
+ "
\n",
+ " \n",
+ " | 965 | \n",
+ " 0.035700 | \n",
+ "
\n",
+ " \n",
+ " | 970 | \n",
+ " 0.029600 | \n",
+ "
\n",
+ " \n",
+ " | 975 | \n",
+ " 0.020600 | \n",
+ "
\n",
+ " \n",
+ " | 980 | \n",
+ " 0.032200 | \n",
+ "
\n",
+ " \n",
+ " | 985 | \n",
+ " 0.040200 | \n",
+ "
\n",
+ " \n",
+ " | 990 | \n",
+ " 0.015200 | \n",
+ "
\n",
+ " \n",
+ " | 995 | \n",
+ " 0.025100 | \n",
+ "
\n",
+ " \n",
+ " | 1000 | \n",
+ " 0.027800 | \n",
+ "
\n",
+ " \n",
+ " | 1005 | \n",
+ " 0.032900 | \n",
+ "
\n",
+ " \n",
+ " | 1010 | \n",
+ " 0.048000 | \n",
+ "
\n",
+ " \n",
+ " | 1015 | \n",
+ " 0.035200 | \n",
+ "
\n",
+ " \n",
+ " | 1020 | \n",
+ " 0.017700 | \n",
+ "
\n",
+ " \n",
+ " | 1025 | \n",
+ " 0.029700 | \n",
+ "
\n",
+ " \n",
+ " | 1030 | \n",
+ " 0.041900 | \n",
+ "
\n",
+ " \n",
+ " | 1035 | \n",
+ " 0.031100 | \n",
+ "
\n",
+ " \n",
+ " | 1040 | \n",
+ " 0.038100 | \n",
+ "
\n",
+ " \n",
+ " | 1045 | \n",
+ " 0.034800 | \n",
+ "
\n",
+ " \n",
+ " | 1050 | \n",
+ " 0.020000 | \n",
+ "
\n",
+ " \n",
+ " | 1055 | \n",
+ " 0.017500 | \n",
+ "
\n",
+ " \n",
+ " | 1060 | \n",
+ " 0.019300 | \n",
+ "
\n",
+ " \n",
+ " | 1065 | \n",
+ " 0.029900 | \n",
+ "
\n",
+ " \n",
+ " | 1070 | \n",
+ " 0.011500 | \n",
+ "
\n",
+ " \n",
+ " | 1075 | \n",
+ " 0.023900 | \n",
+ "
\n",
+ " \n",
+ " | 1080 | \n",
+ " 0.012700 | \n",
+ "
\n",
+ " \n",
+ " | 1085 | \n",
+ " 0.012800 | \n",
+ "
\n",
+ " \n",
+ " | 1090 | \n",
+ " 0.025500 | \n",
+ "
\n",
+ " \n",
+ " | 1095 | \n",
+ " 0.018500 | \n",
+ "
\n",
+ " \n",
+ " | 1100 | \n",
+ " 0.007200 | \n",
+ "
\n",
+ " \n",
+ " | 1105 | \n",
+ " 0.028500 | \n",
+ "
\n",
+ " \n",
+ " | 1110 | \n",
+ " 0.024600 | \n",
+ "
\n",
+ " \n",
+ " | 1115 | \n",
+ " 0.015600 | \n",
+ "
\n",
+ " \n",
+ " | 1120 | \n",
+ " 0.020200 | \n",
+ "
\n",
+ " \n",
+ " | 1125 | \n",
+ " 0.010500 | \n",
+ "
\n",
+ " \n",
+ " | 1130 | \n",
+ " 0.023900 | \n",
+ "
\n",
+ " \n",
+ " | 1135 | \n",
+ " 0.020300 | \n",
+ "
\n",
+ " \n",
+ " | 1140 | \n",
+ " 0.031900 | \n",
+ "
\n",
+ " \n",
+ " | 1145 | \n",
+ " 0.023100 | \n",
+ "
\n",
+ " \n",
+ " | 1150 | \n",
+ " 0.006700 | \n",
+ "
\n",
+ " \n",
+ " | 1155 | \n",
+ " 0.016700 | \n",
+ "
\n",
+ " \n",
+ " | 1160 | \n",
+ " 0.020200 | \n",
+ "
\n",
+ " \n",
+ " | 1165 | \n",
+ " 0.023900 | \n",
+ "
\n",
+ " \n",
+ " | 1170 | \n",
+ " 0.014300 | \n",
+ "
\n",
+ " \n",
+ " | 1175 | \n",
+ " 0.017000 | \n",
+ "
\n",
+ " \n",
+ " | 1180 | \n",
+ " 0.034100 | \n",
+ "
\n",
+ " \n",
+ " | 1185 | \n",
+ " 0.034900 | \n",
+ "
\n",
+ " \n",
+ " | 1190 | \n",
+ " 0.020800 | \n",
+ "
\n",
+ " \n",
+ " | 1195 | \n",
+ " 0.016200 | \n",
+ "
\n",
+ " \n",
+ " | 1200 | \n",
+ " 0.013200 | \n",
+ "
\n",
+ " \n",
+ " | 1205 | \n",
+ " 0.015100 | \n",
+ "
\n",
+ " \n",
+ " | 1210 | \n",
+ " 0.013000 | \n",
+ "
\n",
+ " \n",
+ " | 1215 | \n",
+ " 0.015700 | \n",
+ "
\n",
+ " \n",
+ " | 1220 | \n",
+ " 0.006100 | \n",
+ "
\n",
+ " \n",
+ " | 1225 | \n",
+ " 0.011600 | \n",
+ "
\n",
+ " \n",
+ " | 1230 | \n",
+ " 0.016800 | \n",
+ "
\n",
+ " \n",
+ " | 1235 | \n",
+ " 0.015200 | \n",
+ "
\n",
+ " \n",
+ " | 1240 | \n",
+ " 0.013600 | \n",
+ "
\n",
+ " \n",
+ " | 1245 | \n",
+ " 0.012000 | \n",
+ "
\n",
+ " \n",
+ " | 1250 | \n",
+ " 0.017800 | \n",
+ "
\n",
+ " \n",
+ " | 1255 | \n",
+ " 0.018500 | \n",
+ "
\n",
+ " \n",
+ " | 1260 | \n",
+ " 0.010800 | \n",
+ "
\n",
+ " \n",
+ " | 1265 | \n",
+ " 0.012700 | \n",
+ "
\n",
+ " \n",
+ " | 1270 | \n",
+ " 0.008500 | \n",
+ "
\n",
+ " \n",
+ " | 1275 | \n",
+ " 0.015700 | \n",
+ "
\n",
+ " \n",
+ " | 1280 | \n",
+ " 0.016000 | \n",
+ "
\n",
+ " \n",
+ " | 1285 | \n",
+ " 0.012100 | \n",
+ "
\n",
+ " \n",
+ " | 1290 | \n",
+ " 0.019400 | \n",
+ "
\n",
+ " \n",
+ " | 1295 | \n",
+ " 0.018100 | \n",
+ "
\n",
+ " \n",
+ " | 1300 | \n",
+ " 0.009400 | \n",
+ "
\n",
+ " \n",
+ " | 1305 | \n",
+ " 0.026600 | \n",
+ "
\n",
+ " \n",
+ " | 1310 | \n",
+ " 0.006500 | \n",
+ "
\n",
+ " \n",
+ " | 1315 | \n",
+ " 0.010900 | \n",
+ "
\n",
+ " \n",
+ " | 1320 | \n",
+ " 0.026600 | \n",
+ "
\n",
+ " \n",
+ " | 1325 | \n",
+ " 0.021100 | \n",
+ "
\n",
+ " \n",
+ " | 1330 | \n",
+ " 0.012100 | \n",
+ "
\n",
+ " \n",
+ " | 1335 | \n",
+ " 0.014700 | \n",
+ "
\n",
+ " \n",
+ " | 1340 | \n",
+ " 0.018100 | \n",
+ "
\n",
+ " \n",
+ " | 1345 | \n",
+ " 0.009500 | \n",
+ "
\n",
+ " \n",
+ " | 1350 | \n",
+ " 0.008000 | \n",
+ "
\n",
+ " \n",
+ " | 1355 | \n",
+ " 0.006400 | \n",
+ "
\n",
+ " \n",
+ " | 1360 | \n",
+ " 0.007300 | \n",
+ "
\n",
+ " \n",
+ " | 1365 | \n",
+ " 0.007600 | \n",
+ "
\n",
+ " \n",
+ " | 1370 | \n",
+ " 0.010000 | \n",
+ "
\n",
+ " \n",
+ " | 1375 | \n",
+ " 0.010200 | \n",
+ "
\n",
+ " \n",
+ " | 1380 | \n",
+ " 0.013900 | \n",
+ "
\n",
+ " \n",
+ " | 1385 | \n",
+ " 0.006400 | \n",
+ "
\n",
+ " \n",
+ " | 1390 | \n",
+ " 0.005500 | \n",
+ "
\n",
+ " \n",
+ " | 1395 | \n",
+ " 0.009000 | \n",
+ "
\n",
+ " \n",
+ " | 1400 | \n",
+ " 0.010400 | \n",
+ "
\n",
+ " \n",
+ " | 1405 | \n",
+ " 0.011200 | \n",
+ "
\n",
+ " \n",
+ " | 1410 | \n",
+ " 0.007200 | \n",
+ "
\n",
+ " \n",
+ " | 1415 | \n",
+ " 0.008300 | \n",
+ "
\n",
+ " \n",
+ " | 1420 | \n",
+ " 0.006000 | \n",
+ "
\n",
+ " \n",
+ " | 1425 | \n",
+ " 0.010000 | \n",
+ "
\n",
+ " \n",
+ " | 1430 | \n",
+ " 0.008900 | \n",
+ "
\n",
+ " \n",
+ " | 1435 | \n",
+ " 0.013300 | \n",
+ "
\n",
+ " \n",
+ " | 1440 | \n",
+ " 0.015000 | \n",
+ "
\n",
+ " \n",
+ " | 1445 | \n",
+ " 0.013900 | \n",
+ "
\n",
+ " \n",
+ " | 1450 | \n",
+ " 0.016000 | \n",
+ "
\n",
+ " \n",
+ " | 1455 | \n",
+ " 0.019200 | \n",
+ "
\n",
+ " \n",
+ " | 1460 | \n",
+ " 0.012300 | \n",
+ "
\n",
+ " \n",
+ " | 1465 | \n",
+ " 0.012800 | \n",
+ "
\n",
+ " \n",
+ " | 1470 | \n",
+ " 0.011900 | \n",
+ "
\n",
+ " \n",
+ " | 1475 | \n",
+ " 0.016400 | \n",
+ "
\n",
+ " \n",
+ " | 1480 | \n",
+ " 0.008100 | \n",
+ "
\n",
+ " \n",
+ " | 1485 | \n",
+ " 0.007700 | \n",
+ "
\n",
+ " \n",
+ " | 1490 | \n",
+ " 0.006300 | \n",
+ "
\n",
+ " \n",
+ " | 1495 | \n",
+ " 0.002100 | \n",
+ "
\n",
+ " \n",
+ " | 1500 | \n",
+ " 0.008000 | \n",
+ "
\n",
+ " \n",
+ "
"
+ ],
+ "text/plain": [
+ ""
+ ]
+ },
+ "metadata": {},
+ "output_type": "display_data"
+ }
+ ],
+ "source": [
+ "#trainer_stats = trainer.train()\n",
+ "trainer_stats = trainer.train(resume_from_checkpoint = False)\n"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 18,
+ "metadata": {},
+ "outputs": [
+ {
+ "name": "stdout",
+ "output_type": "stream",
+ "text": [
+ "llamafactory支持多种类型的微调，具体支持的类型可以通过配置文件的`finetuning_type`字段来指定。例如，在`examples/inference/llamafactory_lora_sft.yaml`配置文件中，`finetuning_type`指定为`lora`，表示使用LoRA微调方法。<|im_end|>\n"
+ ]
+ }
+ ],
+ "source": [
+ "FastLanguageModel.for_inference(model)\n",
+ "\n",
+ "# 准备相同的测试输入\n",
+ "messages = [\n",
+ " {\"role\": \"user\", \"content\": \n",
+ " \"\"\"llama factory支持什么类型的微调\"\"\"}\n",
+ "]\n",
+ "inputs = tokenizer.apply_chat_template(\n",
+ " messages,\n",
+ " tokenize = True,\n",
+ " add_generation_prompt = True,\n",
+ " return_tensors = \"pt\",\n",
+ ").to(\"cuda\")\n",
+ "\n",
+ "# 使用TextStreamer进行流式生成\n",
+ "from transformers import TextStreamer\n",
+ "text_streamer = TextStreamer(tokenizer, skip_prompt = True)\n",
+ "_ = model.generate(input_ids = inputs, streamer = text_streamer, max_new_tokens = 128,\n",
+ " use_cache = True, temperature = 1.5, min_p = 0.1)"
+ ]
+ }
+ ],
+ "metadata": {
+ "kernelspec": {
+ "display_name": "Python 3",
+ "language": "python",
+ "name": "python3"
+ },
+ "language_info": {
+ "codemirror_mode": {
+ "name": "ipython",
+ "version": 3
+ },
+ "file_extension": ".py",
+ "mimetype": "text/x-python",
+ "name": "python",
+ "nbconvert_exporter": "python",
+ "pygments_lexer": "ipython3",
+ "version": "3.11.11"
+ }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}
diff --git a/trainer.py b/trainer.py
new file mode 100644
index 0000000..740acb2
--- /dev/null
+++ b/trainer.py
@@ -0,0 +1,70 @@
+from unsloth import FastLanguageModel
+import torch
+
+# 基础配置参数
+max_seq_length = 4096 # 最大序列长度
+dtype = None # 自动检测数据类型
+load_in_4bit = True # 使用4位量化以减少内存使用
+
+# 加载预训练模型和分词器
+model, tokenizer = FastLanguageModel.from_pretrained(
+ model_name = "workdir\model\Qwen2.5-3B-Instruct-bnb-4bit", # 选择Qwen2.5 32B指令模型
+ max_seq_length = max_seq_length,
+ dtype = dtype,
+ load_in_4bit = load_in_4bit,
+)
+
+model = FastLanguageModel.get_peft_model(
+ model,
+ r = 64, # LoRA秩,控制可训练参数数量
+ target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
+ "gate_proj", "up_proj", "down_proj",], # 需要训练的目标模块
+ lora_alpha = 64, # LoRA缩放因子
+ lora_dropout = 0, # LoRA dropout率
+ bias = "none", # 是否训练偏置项
+ use_gradient_checkpointing = "unsloth", # 使用梯度检查点节省显存
+ random_state = 114514, # 随机数种子
+ use_rslora = False, # 是否使用稳定版LoRA
+ loftq_config = None, # LoftQ配置
+)
+
+from unsloth.chat_templates import get_chat_template
+# 配置分词器使用qwen-2.5对话模板
+tokenizer = get_chat_template(
+ tokenizer,
+ chat_template="qwen-2.5",
+)
+
+def formatting_prompts_func(examples):
+ """格式化对话数据的函数
+ Args:
+ examples: 包含对话列表的字典
+ Returns:
+ 包含格式化文本的字典
+ """
+ questions = examples["question"]
+ answer = examples["answer"]
+
+ # 将Question和Response组合成对话形式
+ convos = [
+ [{"role": "user", "content": q}, {"role": "assistant", "content": r}]
+ for q, r in zip(questions, answer)
+ ]
+
+ # 使用tokenizer.apply_chat_template格式化对话
+ texts = [
+ tokenizer.apply_chat_template(convo, tokenize=False, add_generation_prompt=False)
+ for convo in convos
+ ]
+
+ return {"text": texts}
+
+from unsloth.chat_templates import standardize_sharegpt
+
+# 加载数据集
+from datasets import load_dataset
+dataset = load_dataset("json", data_files="workdir\dataset\dataset.json")
+dataset = dataset.map(formatting_prompts_func, batched = True)
+
+print(dataset[5])
+print(dataset[5]["text"])
\ No newline at end of file