一、核心工具
模型微调全流程需安装以下工具：
- 必装工具：Unsloth（高效微调框架）
- 可选工具：
- vLLM（模型调度与推理验证）
- EvalScope（模型性能评估）
- wandb（训练过程监控）
若追求快速上手，可先安装 Unsloth；若需完整流程验证，建议全量安装。
二、Unsloth 安装
Unsloth 是专为大模型设计的动态量化与微调框架，能显著提升微调速度并降低显存占用。
虚拟环境创建
# 创建虚拟环境 conda create --name unsloth python=3.11 conda init source ~/.bashrc conda activate unsloth # 安装Jupyter相关组件 conda install jupyterlab conda install ipykernel python -m ipykernel install --user --name unsloth --display-name "Python unsloth"框架安装
pip install --upgrade --force-reinstall --no-cache-dir unsloth unsloth_zoo安装验证
在 Jupyter 中选择 “Python unsloth” 内核，运行以下代码测试：
from unsloth import FastLanguageModel import torch无报错则安装成功。
核心优势
- 微调速度提升 2-5 倍，显存使用减少 80%
- 兼容 HuggingFace 生态，支持 SFT 和 DPO 等微调方式
- 支持 4 位动态量化，在低显存设备上也能运行大模型
三、vLLM 安装
vLLM 是高效的模型推理框架，用于微调后模型效果验证。
虚拟环境配置
# 创建独立虚拟环境（AutoDL环境可跳过） conda create --name vllm python=3.11 conda init source ~/.bashrc conda activate vllm框架安装
# 安装依赖库 pip install bitsandbytes>=0.45.3 # 安装vLLM pip install --upgrade vllm安装模型下载工具
pip install modelscope下载模型权重
以 32B 4bit 动态量化模型为例：
modelscope download --model unsloth/Qwen3-32B-unsloth-bnb-4bit --local_dir ./Qwen3-32B-unsloth-bnb-4bit模型调用测试
启动vLLM 服务
vllm serve ./Qwen3-32B-unsloth-bnb-4bit --enable-auto-tool-choice --tool-call-parser hermesvLLM调用qwen3参数说明，详见：https://qwen.readthedocs.io/en/latest/deployment/vllm.html#
代码调用测试
from openai import OpenAI openai_api_key = "EMPTY" openai_api_base = "http://localhost:8000/v1" client = OpenAI( api_key=openai_api_key, base_url=openai_api_base, ) messages = [{"role": "user", "content": "你好,好久不见!"}] response = client.chat.completions.create( model="./Qwen3-32B-unsloth-bnb-4bit", messages=messages, ) print(response.choices[0].message.content)注意
- Unsloth 动态量化模型仅支持单卡调用
- 32B 4bit 模型最低需 22G 显存调用，37G 显存微调
- 目前 vLLM 仅支持 dense 模型，MoE 模型暂不兼容
四、EvalScope 安装部署流程
EvalScope 是大模型性能评估框架，用于对比微调前后模型性能变化。
4.1 环境搭建
# 创建独立虚拟环境 conda create --name evalscope python=3.11 conda init source ~/.bashrc conda activate evalscope # 安装Jupyter组件 conda install jupyterlab conda install ipykernel python -m ipykernel install --user --name evalscope --display-name "Python evalscope"4.2 安装
# 基础安装 pip install evalscope # 可选扩展（根据需求安装） pip install 'evalscope[opencompass]' # OpenCompass后端 pip install 'evalscope[vlmeval]' # VLMEvalKit后端 pip install 'evalscope[rag]' # RAGEval后端 pip install 'evalscope[all]' # 安装所有后端4.3 压力测试示例
evalscope perf \ --url "http://127.0.0.1:8000/v1/chat/completions" \ --parallel 5 \ --model ./Qwen3-32B-unsloth-bnb-4bit \ --number 20 \ --api openai \ --dataset openqa \ --stream4.4 模型性能评估
这里采用了EvalScope专为Qwen3准备的 modelscope/EvalScope-Qwen3-Test 数据集进行评测，会
围绕模型的推理、指令跟随、代理能力和多语言支持方面能力进行测试，该数据包含 mmlu_pro 、
ifeval 、 live_code_bench 、 math_500 、 aime24 等各著名评估数据集。
数据集地址：https://modelscope.cn/datasets/modelscope/EvalScope-Qwen3-Test/summary
估代码如下：
from evalscope import TaskConfig, run_task task_cfg = TaskConfig( model='./Qwen3-32B-unsloth-bnb-4bit', api_url='http://127.0.0.1:8000/v1/chat/completions', eval_type='service', datasets=['data_collection'], dataset_args={ 'data_collection': { 'dataset_id': 'modelscope/EvalScope-Qwen3-Test', 'filters': {'remove_until': '</think>'} } }, eval_batch_size=128, generation_config={ 'max_tokens': 30000, 'temperature': 0.6, 'top_p': 0.95, 'top_k': 20, 'n': 1, }, timeout=60000, stream=True, limit=2000, ) run_task(task_cfg=task_cfg)4.5评估结果可视化
evalscope app访问输出的本地 URL 即可查看评估报告。
五、wandb 安装与注册
wandb 是模型训练可视化工具，用于监控训练过程中的关键指标。
安装
pip install wandb注册与登录
- 访问wandb 官网(https://wandb.ai/site)注册账号
- 在终端运行登录命令，按提示输入 API 密钥：
wandb login核心功能
- 实时可视化训练指标（损失值、学习率等）
- 自动记录实验参数与结果，确保可追溯性
- 支持训练中断与恢复
- 多实验对比分析，辅助调优决策
使用方法
在微调代码中加入 wandb 初始化代码：
import wandb wandb.init(project="qwen3-finetuning", config={"epochs": 3, "learning_rate": 2e-4})六、相关代码
没仔细整理，做个记录吧，代码来源于网络，侵删。
from unsloth import FastLanguageModel import torch import requests, json import pandas as pd from datasets import load_dataset from datasets import load_from_disk from unsloth.chat_templates import standardize_sharegpt from trl import SFTTrainer, SFTConfig import wandb import os os.environ["CUDA_VISIBLE_DEVICES"] = "1" max_seq_length = 8192 dtype = None load_in_4bit = True model, tokenizer = FastLanguageModel.from_pretrained( model_name = "./Qwen3-32B-unsloth-bnb-4bit", max_seq_length = max_seq_length, dtype = dtype, load_in_4bit = load_in_4bit, ) print(model) print(tokenizer) gpu_stats = torch.cuda.get_device_properties(0) start_gpu_memory = round(torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024, 3) max_memory = round(gpu_stats.total_memory / 1024 / 1024 / 1024, 3) print(f"GPU = {gpu_stats.name}. Max memory = {max_memory} GB.") print(f"{start_gpu_memory} GB of memory reserved.") messages = [ {"role" : "user", "content" : "你好，好久不见！"} ] # =============不设置思考=============== text = tokenizer.apply_chat_template( messages, tokenize = False, add_generation_prompt = True, enable_thinking = False, # 设置不思考 ) # '<|im_start|>user\n你好，好久不见！<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n' print(text) inputs = tokenizer(text, return_tensors="pt").to("cuda") outputs = model.generate( input_ids=inputs.input_ids, attention_mask=inputs.attention_mask, max_new_tokens=max_seq_length, use_cache=True, ) print(outputs) response = tokenizer.batch_decode(outputs) print(response) # =============设置思考=============== text = tokenizer.apply_chat_template( messages, tokenize = False, add_generation_prompt = True, enable_thinking = True, # 设置思考 ) inputs = tokenizer(text, return_tensors="pt").to("cuda") outputs = model.generate( input_ids=inputs.input_ids, attention_mask=inputs.attention_mask, max_new_tokens=max_seq_length, use_cache=True, ) response = tokenizer.batch_decode(outputs) messages = [ {"role" : "system", "content" : "你是一名助人为乐的助手，名叫小明。"}, {"role" : "user", "content" : "你好，好久不见！请问你叫什么名字？"} ] text = tokenizer.apply_chat_template( messages, tokenize = False, add_generation_prompt = True, enable_thinking = True, # 设置思考 ) inputs = tokenizer(text, return_tensors="pt").to("cuda") outputs = model.generate( input_ids=inputs.input_ids, attention_mask=inputs.attention_mask, max_new_tokens=max_seq_length, use_cache=True, ) response = tokenizer.batch_decode(outputs) # ============= function calling =============== def get_weather(loc): """ 查询即时天气函数 :param loc: 必要参数，字符串类型，用于表示查询天气的具体城市名称，\ 注意，中国的城市需要用对应城市的英文名称代替，例如如果需要查询北京市天气，则loc参数需要输入'Beijing'； :return：OpenWeather API查询即时天气的结果，具体URL请求地址为：https://api.openweathermap.org/data/2.5/weather\ 返回结果对象类型为解析之后的JSON格式对象，并用字符串形式进行表示，其中包含了全部重要的天气信息 """ # Step 1.构建请求 url = "https://api.openweathermap.org/data/2.5/weather" # Step 2.设置查询参数 params = { "q": loc, "appid": "YOUR_API_KEY", # 输入API key "units": "metric", # 使用摄氏度而不是华氏度 "lang":"zh_cn" # 输出语言为简体中文 } # Step 3.发送GET请求 response = requests.get(url, params=params) # Step 4.解析响应 data = response.json() return json.dumps(data) tools = [ { "type": "function", "function":{ 'name': 'get_weather', 'description': '查询即时天气函数，根据输入的城市名称，查询对应城市的实时天气，一次只能输入一个城市名称', 'parameters': { 'type': 'object', 'properties': { 'loc': { 'description': "城市名称，注意，中国的城市需要用对应城市的英文名称代替，例如如果需要查询北京市天气，则loc参数需要输入'Beijing'", 'type': 'string' } }, 'required': ['loc'] } } } ] messages = [ {"role" : "system", "content" : "你是一名助人为乐的天气查询助手，当用户询问天气信息时，请调用get_weather函数进行天气查询。"}, {"role" : "user", "content" : "你好，请帮我查询下北京今天天气如何？"} ] text = tokenizer.apply_chat_template( messages, tools = tools, tokenize = False, add_generation_prompt = True, enable_thinking = True, # 设置思考 ) inputs = tokenizer(text, return_tensors="pt").to("cuda") outputs = model.generate( input_ids=inputs.input_ids, attention_mask=inputs.attention_mask, max_new_tokens=max_seq_length, use_cache=True, ) response = tokenizer.batch_decode(outputs) messages = [ {"role" : "system", "content" : "你是一名助人为乐的天气查询助手，当用户询问天气信息时，请调用get_weather函数进行天气查询。"}, {"role" : "user", "content" : "你好，请帮我查询下北京和杭州今天天气如何？"} ] text = tokenizer.apply_chat_template( messages, tools = tools, tokenize = False, add_generation_prompt = True, enable_thinking = True, # 设置思考 ) inputs = tokenizer(text, return_tensors="pt").to("cuda") outputs = model.generate( input_ids=inputs.input_ids, attention_mask=inputs.attention_mask, max_new_tokens=max_seq_length, use_cache=True, ) response = tokenizer.batch_decode(outputs) messages.append({ "role": "assistant", "content": "<think>\n我将调用 get_weather 函数来查询天气。\n</think>\n", "tool_calls": [ { "name": "get_weather", "arguments": { "location": "北京" } }, { "name": "get_weather", "arguments": { "location": "杭州" } } ] }) messages.append({ "role": "tool", "content": json.dumps({ "location": "北京", "weather": "晴，最高气温26℃" }) }) messages.append({ "role": "tool", "content": json.dumps({ "location": "杭州", "weather": "多云转小雨，最高气温23℃" }) }) text = tokenizer.apply_chat_template( messages, tools = tools, tokenize = False, add_generation_prompt = True, enable_thinking = True, # 设置思考 ) inputs = tokenizer(text, return_tensors="pt").to("cuda") outputs = model.generate( input_ids=inputs.input_ids, attention_mask=inputs.attention_mask, max_new_tokens=max_seq_length, use_cache=True, ) response = tokenizer.batch_decode(outputs) # ============= stream 模式 =============== from transformers import TextStreamer messages = [ {"role" : "user", "content" : "你好，好久不见！"} ] text = tokenizer.apply_chat_template( messages, tokenize = False, add_generation_prompt = True, enable_thinking = False, ) _ = model.generate( **tokenizer(text, return_tensors = "pt").to("cuda"), max_new_tokens = 256, # Increase for longer outputs! temperature = 0.7, top_p = 0.8, top_k = 20, # For non thinking streamer = TextStreamer(tokenizer, skip_prompt = True), ) text = tokenizer.apply_chat_template( messages, tokenize = False, add_generation_prompt = True, enable_thinking = True, ) _ = model.generate( **tokenizer(text, return_tensors = "pt").to("cuda"), max_new_tokens = 2048, # Increase for longer outputs! temperature = 0.6, top_p = 0.95, top_k = 20, # For thinking streamer = TextStreamer(tokenizer, skip_prompt = True), ) reasoning_dataset = load_dataset("unsloth/OpenMathReasoning-mini", split = "cot") non_reasoning_dataset = load_dataset("mlabonne/FineTome-100k", split = "train") print(len(reasoning_dataset), reasoning_dataset[0]) def generate_conversation(examples): problems = examples["problem"] solutions = examples["generated_solution"] conversations = [] for problem, solution in zip(problems, solutions): conversations.append([ {"role" : "user", "content" : problem}, {"role" : "assistant", "content" : solution}, ]) return { "conversations": conversations, } reasoning_data = reasoning_dataset.map(generate_conversation, batched = True) print(reasoning_data["conversations"][0]) reasoning_conversations = tokenizer.apply_chat_template( reasoning_data["conversations"], tokenize = False, ) print(len(reasoning_conversations), reasoning_conversations[0]) dataset = standardize_sharegpt(non_reasoning_dataset) print(dataset["conversations"][0]) non_reasoning_conversations = tokenizer.apply_chat_template( dataset["conversations"], tokenize = False, ) print(non_reasoning_conversations[0]) print(len(reasoning_conversations)) print(len(non_reasoning_conversations)) chat_percentage = 0.75 non_reasoning_subset = pd.Series(non_reasoning_conversations) non_reasoning_subset = non_reasoning_subset.sample( int(len(reasoning_conversations) * (1.0 - chat_percentage)), random_state = 2407, ) data = pd.concat([ pd.Series(reasoning_conversations), pd.Series(non_reasoning_subset) ]) data.name = "text" from datasets import Dataset combined_dataset = Dataset.from_pandas(pd.DataFrame(data)) combined_dataset = combined_dataset.shuffle(seed = 3407) print(len(combined_dataset), type(combined_dataset)) combined_dataset.save_to_disk("cleaned_qwen3_dataset") combined_dataset = load_from_disk("cleaned_qwen3_dataset") print(type(combined_dataset)) model = FastLanguageModel.get_peft_model( model, r = 32, # Choose any number > 0! Suggested 8, 16, 32, 64, 128 target_modules = ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj",], lora_alpha = 32, # Best to choose alpha = rank or rank*2 lora_dropout = 0, # Supports any, but = 0 is optimized bias = "none", # Supports any, but = "none" is optimized # [NEW] "unsloth" uses 30% less VRAM, fits 2x larger batch sizes! use_gradient_checkpointing = "unsloth", # True or "unsloth" for very long context random_state = 3407, use_rslora = False, # We support rank stabilized LoRA loftq_config = None, # And LoftQ ) trainer = SFTTrainer( model = model, tokenizer = tokenizer, train_dataset = combined_dataset, eval_dataset = None, # Can set up evaluation! args = SFTConfig( dataset_text_field = "text", per_device_train_batch_size = 4, gradient_accumulation_steps = 2, # Use GA to mimic batch size! warmup_steps = 5, num_train_epochs = 1, # Set this for 1 full training run. learning_rate = 2e-4, # Reduce to 2e-5 for long training runs logging_steps = 1, optim = "adamw_8bit", weight_decay = 0.01, lr_scheduler_type = "linear", seed = 3407, report_to = "none", # Use this for WandB etc ), ) # @title Show current memory stats gpu_stats = torch.cuda.get_device_properties(0) start_gpu_memory = round(torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024, 3) max_memory = round(gpu_stats.total_memory / 1024 / 1024 / 1024, 3) print(f"GPU = {gpu_stats.name}. Max memory = {max_memory} GB.") print(f"{start_gpu_memory} GB of memory reserved.") os.environ["WANDB_NOTEBOOK_NAME"] = "notebook_name.ipynb" wandb.login(key="your_key") run = wandb.init(project='Fine-tune-Qwen-32B-4bit on Combined Dataset', ) trainer_stats = trainer.train() # @title Show final memory and time stats used_memory = round(torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024, 3) used_memory_for_lora = round(used_memory - start_gpu_memory, 3) used_percentage = round(used_memory / max_memory * 100, 3) lora_percentage = round(used_memory_for_lora / max_memory * 100, 3) print(f"{trainer_stats.metrics['train_runtime']} seconds used for training.") print( f"{round(trainer_stats.metrics['train_runtime']/60, 2)} minutes used for training." ) print(f"Peak reserved memory = {used_memory} GB.") print(f"Peak reserved memory for training = {used_memory_for_lora} GB.") print(f"Peak reserved memory % of max memory = {used_percentage} %.") print(f"Peak reserved memory for training % of max memory = {lora_percentage} %.") print(model) messages = [ {"role" : "user", "content" : "Determine the surface area of the portion of the plane $2x + 3y + 6z = 9$ that lies in the first octant."} ] text = tokenizer.apply_chat_template( messages, tokenize = False, add_generation_prompt = True, # Must add for generation enable_thinking = True, # Disable thinking ) _ = model.generate( **tokenizer(text, return_tensors = "pt").to("cuda"), max_new_tokens = 20488, # Increase for longer outputs! temperature = 0.6, top_p = 0.95, top_k = 20, # For thinking streamer = TextStreamer(tokenizer, skip_prompt = True), ) model.save_pretrained_merged(save_directory = "Qwen3-32B-finetuned-fp16", tokenizer = tokenizer, save_method = "merged_16bit") max_seq_length = 8192 dtype = None load_in_4bit = True model, tokenizer = FastLanguageModel.from_pretrained( model_name = "./Qwen3-8B-unsloth-bnb-4bit", max_seq_length = max_seq_length, dtype = dtype, load_in_4bit = load_in_4bit, ) question_1 = "中国在哪里" messages = [ {"role" : "user", "content" : question_1} ] text = tokenizer.apply_chat_template( messages, tokenize = False, add_generation_prompt = True, # Must add for generation enable_thinking = True, # Disable thinking ) _ = model.generate( **tokenizer(text, return_tensors = "pt").to("cuda"), max_new_tokens = 20488, # Increase for longer outputs! temperature = 0.6, top_p = 0.95, top_k = 20, # For thinking streamer = TextStreamer(tokenizer, skip_prompt = True), ) def load_jsonl_dataset(file_path): """从JSONL文件加载数据集""" data = {"input": [], "output": []} print(f"开始加载数据集: {file_path}") count = 0 error_count = 0 with open(file_path, 'r', encoding='utf-8') as f: for line in f: try: item = json.loads(line.strip()) # 根据数据集结构提取字段 input_text = item.get("input", "") output = item.get("output", "") data["input"].append(input_text) data["output"].append(output) count += 1 except Exception as e: print(f"解析行时出错: {e}") error_count += 1 continue print(f"数据集加载完成: 成功加载{count}个样本, 跳过{error_count}个错误样本") return Dataset.from_dict(data) data_path = "./train_1k.jsonl" # 加载自定义数据集 dataset = load_jsonl_dataset(data_path) # 显示数据集信息 print(f"\n数据集统计:") print(f"- 样本数量: {len(dataset)}") print(f"- 字段: {dataset.column_names}") print(dataset[0]) def formatting_prompts_func(examples): """根据提示模板格式化数据""" inputs = examples["input"] outputs = examples["output"] texts = [] for input_text, output in zip(inputs, outputs): texts.append([ {"role" : "user", "content" : input_text}, {"role" : "assistant", "content" : output}, ]) return {"text": texts} # 应用格式化 print("开始格式化数据集...") reasoning_conversations = tokenizer.apply_chat_template( dataset.map(formatting_prompts_func, batched = True)["text"], tokenize = False, ) print("数据集格式化完成") non_reasoning_dataset = load_dataset("mlabonne/FineTome-100k", split = "train") dataset = standardize_sharegpt(non_reasoning_dataset) non_reasoning_conversations = tokenizer.apply_chat_template( dataset["conversations"], tokenize = False, ) non_reasoning_subset = pd.Series(non_reasoning_conversations) non_reasoning_subset = non_reasoning_subset.sample( 1000, random_state = 2407, ) data = pd.concat([ pd.Series(reasoning_conversations), pd.Series(non_reasoning_subset) ]) data.name = "text" combined_dataset = Dataset.from_pandas(pd.DataFrame(data)) combined_dataset = combined_dataset.shuffle(seed = 3407) model = FastLanguageModel.get_peft_model( model, r = 32, # Choose any number > 0! Suggested 8, 16, 32, 64, 128 target_modules = ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj",], lora_alpha = 32, # Best to choose alpha = rank or rank*2 lora_dropout = 0, # Supports any, but = 0 is optimized bias = "none", # Supports any, but = "none" is optimized # [NEW] "unsloth" uses 30% less VRAM, fits 2x larger batch sizes! use_gradient_checkpointing = "unsloth", # True or "unsloth" for very long context random_state = 3407, use_rslora = False, # We support rank stabilized LoRA loftq_config = None, # And LoftQ ) trainer = SFTTrainer( model = model, tokenizer = tokenizer, train_dataset = combined_dataset, eval_dataset = None, # Can set up evaluation! args = SFTConfig( dataset_text_field = "text", per_device_train_batch_size = 2, gradient_accumulation_steps = 4, # Use GA to mimic batch size! warmup_steps = 5, num_train_epochs = 1, # Set this for 1 full training run. learning_rate = 2e-4, # Reduce to 2e-5 for long training runs logging_steps = 1, optim = "adamw_8bit", weight_decay = 0.01, lr_scheduler_type = "linear", seed = 3407, report_to = None, # Use this for WandB etc ), ) # @title Show current memory stats gpu_stats = torch.cuda.get_device_properties(0) start_gpu_memory = round(torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024, 3) max_memory = round(gpu_stats.total_memory / 1024 / 1024 / 1024, 3) print(f"GPU = {gpu_stats.name}. Max memory = {max_memory} GB.") print(f"{start_gpu_memory} GB of memory reserved.") trainer_stats = trainer.train() # @title Show final memory and time stats used_memory = round(torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024, 3) used_memory_for_lora = round(used_memory - start_gpu_memory, 3) used_percentage = round(used_memory / max_memory * 100, 3) lora_percentage = round(used_memory_for_lora / max_memory * 100, 3) print(f"{trainer_stats.metrics['train_runtime']} seconds used for training.") print( f"{round(trainer_stats.metrics['train_runtime']/60, 2)} minutes used for training." ) print(f"Peak reserved memory = {used_memory} GB.") print(f"Peak reserved memory for training = {used_memory_for_lora} GB.") print(f"Peak reserved memory % of max memory = {used_percentage} %.") print(f"Peak reserved memory for training % of max memory = {lora_percentage} %.") messages = [ {"role" : "user", "content" : question_1} ] text = tokenizer.apply_chat_template( messages, tokenize = False, add_generation_prompt = True, # Must add for generation enable_thinking = True, # Disable thinking ) _ = model.generate( **tokenizer(text, return_tensors = "pt").to("cuda"), max_new_tokens = 20488, # Increase for longer outputs! temperature = 0.6, top_p = 0.95, top_k = 20, # For thinking streamer = TextStreamer(tokenizer, skip_prompt = True), )想入门 AI 大模型却找不到清晰方向？备考大厂 AI 岗还在四处搜集零散资料？别再浪费时间啦！2025 年AI 大模型全套学习资料已整理完毕，从学习路线到面试真题，从工具教程到行业报告，一站式覆盖你的所有需求，现在全部免费分享！
👇👇扫码免费领取全部内容👇👇
一、学习必备：100+本大模型电子书+26 份行业报告 + 600+ 套技术PPT，帮你看透 AI 趋势
想了解大模型的行业动态、商业落地案例？大模型电子书？这份资料帮你站在 “行业高度” 学 AI：
1. 100+本大模型方向电子书
2. 26 份行业研究报告：覆盖多领域实践与趋势
报告包含阿里、DeepSeek 等权威机构发布的核心内容，涵盖：
- 职业趋势：《AI + 职业趋势报告》《中国 AI 人才粮仓模型解析》；
- 商业落地：《生成式 AI 商业落地白皮书》《AI Agent 应用落地技术白皮书》；
- 领域细分：《AGI 在金融领域的应用报告》《AI GC 实践案例集》；
- 行业监测：《2024 年中国大模型季度监测报告》《2025 年中国技术市场发展趋势》。
3. 600+套技术大会 PPT：听行业大咖讲实战
PPT 整理自 2024-2025 年热门技术大会，包含百度、腾讯、字节等企业的一线实践：
- 安全方向：《端侧大模型的安全建设》《大模型驱动安全升级（腾讯代码安全实践）》；
- 产品与创新：《大模型产品如何创新与创收》《AI 时代的新范式：构建 AI 产品》；
- 多模态与 Agent：《Step-Video 开源模型（视频生成进展）》《Agentic RAG 的现在与未来》；
- 工程落地：《从原型到生产：AgentOps 加速字节 AI 应用落地》《智能代码助手 CodeFuse 的架构设计》。
二、求职必看：大厂 AI 岗面试 “弹药库”，300 + 真题 + 107 道面经直接抱走
想冲字节、腾讯、阿里、蔚来等大厂 AI 岗？这份面试资料帮你提前 “押题”，拒绝临场慌！
1. 107 道大厂面经：覆盖 Prompt、RAG、大模型应用工程师等热门岗位
面经整理自 2021-2025 年真实面试场景，包含 TPlink、字节、腾讯、蔚来、虾皮、中兴、科大讯飞、京东等企业的高频考题，每道题都附带思路解析：
2. 102 道 AI 大模型真题：直击大模型核心考点
针对大模型专属考题，从概念到实践全面覆盖，帮你理清底层逻辑：
3. 97 道 LLMs 真题：聚焦大型语言模型高频问题
专门拆解 LLMs 的核心痛点与解决方案，比如让很多人头疼的 “复读机问题”：
![]()
三、路线必明： AI 大模型学习路线图，1 张图理清核心内容
刚接触 AI 大模型，不知道该从哪学起？这份「AI大模型 学习路线图」直接帮你划重点，不用再盲目摸索！
路线图涵盖 5 大核心板块，从基础到进阶层层递进：一步步带你从入门到进阶，从理论到实战。
L1阶段:启航篇丨极速破界AI新时代
L1阶段：了解大模型的基础知识，以及大模型在各个行业的应用和分析，学习理解大模型的核心原理、关键技术以及大模型应用场景。
L2阶段：攻坚篇丨RAG开发实战工坊
L2阶段：AI大模型RAG应用开发工程，主要学习RAG检索增强生成：包括Naive RAG、Advanced-RAG以及RAG性能评估，还有GraphRAG在内的多个RAG热门项目的分析。
L3阶段：跃迁篇丨Agent智能体架构设计
L3阶段：大模型Agent应用架构进阶实现，主要学习LangChain、 LIamaIndex框架，也会学习到AutoGPT、 MetaGPT等多Agent系统，打造Agent智能体。
L4阶段：精进篇丨模型微调与私有化部署
L4阶段：大模型的微调和私有化部署，更加深入的探讨Transformer架构，学习大模型的微调技术，利用DeepSpeed、Lamam Factory等工具快速进行模型微调，并通过Ollama、vLLM等推理部署框架，实现模型的快速部署。
L5阶段：专题集丨特训篇 【录播课】
![]()
四、资料领取：全套内容免费抱走，学 AI 不用再找第二份
不管你是 0 基础想入门 AI 大模型，还是有基础想冲刺大厂、了解行业趋势，这份资料都能满足你！
现在只需按照提示操作，就能免费领取：
👇👇扫码免费领取全部内容👇👇
2025 年想抓住 AI 大模型的风口？别犹豫，这份免费资料就是你的 “起跑线”！
)
