Upload 4 files

app.py

@@ -0,0 +1,37 @@
+import subprocess
+import gradio as gr
+
+def train():
+    # 安装依赖
+    process = subprocess.Popen(
+        ['pip', 'install', '-r', 'requirements.txt'],
+        stdout=subprocess.PIPE,
+        stderr=subprocess.STDOUT,
+        text=True
+    )
+    for line in iter(process.stdout.readline, ''):
+        yield line
+    process.wait()
+    
+    yield "---依赖安装完成，开始训练---"
+
+    # 运行训练脚本
+    process = subprocess.Popen(
+        ['python3', 'fine_tune.py'],
+        stdout=subprocess.PIPE,
+        stderr=subprocess.STDOUT,
+        text=True
+    )
+    for line in iter(process.stdout.readline, ''):
+        yield line
+    process.wait()
+    
+    yield "---训练完成！---"
+
+with gr.Blocks() as demo:
+    gr.Markdown("点击按钮开始微调")
+    output = gr.Textbox(label="训练日志", lines=20)
+    train_button = gr.Button("开始微调")
+    train_button.click(fn=train, inputs=[], outputs=output)
+
+demo.launch()

data.json

@@ -0,0 +1,22 @@
+[
+    {
+        "instruction": "根据以下信息，生成一个用户JSON对象。",
+        "input": "用户ID是123，用户名是alice，邮箱是alice@example.com",
+        "output": "{\"user_id\": 123, \"username\": \"alice\", \"email\": \"alice@example.com\"}"
+    },
+    {
+        "instruction": "根据以下信息，生成一个用户JSON对象。",
+        "input": "用户ID是456，用户名是bob，邮箱是bob@example.com",
+        "output": "{\"user_id\": 456, \"username\": \"bob\", \"email\": \"bob@example.com\"}"
+    },
+    {
+        "instruction": "根据以下信息，生成一个用户JSON对象。",
+        "input": "用户ID是789，用户名是charlie，邮箱是charlie@example.com",
+        "output": "{\"user_id\": 789, \"username\": \"charlie\", \"email\": \"charlie@example.com\"}"
+    },
+    {
+        "instruction": "根据以下信息，生成一个用户JSON对象。",
+        "input": "用户ID是101，用户名是dave，邮箱是dave@example.com",
+        "output": "{\"user_id\": 101, \"username\": \"dave\", \"email\": \"dave@example.com\"}"
+    }
+]

fine_tune.py

@@ -0,0 +1,82 @@
+import torch
+from datasets import load_dataset
+from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training
+from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, TrainingArguments
+from trl import SFTTrainer
+
+# 1. 加载模型和分词器
+model_name = "NousResearch/Llama-2-7b-chat-hf"
+
+# BitsAndBytesConfig for QLoRA
+bnb_config = BitsAndBytesConfig(
+    load_in_4bit=True,
+    bnb_4bit_use_double_quant=True,
+    bnb_4bit_quant_type="nf4",
+    bnb_4bit_compute_dtype=torch.bfloat16
+)
+
+# Load model
+model = AutoModelForCausalLM.from_pretrained(
+    model_name,
+    quantization_config=bnb_config,
+    device_map="auto",
+    trust_remote_code=True
+)
+model.config.use_cache = False # for training
+
+tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
+tokenizer.pad_token = tokenizer.eos_token # set pad token
+
+# 2. 加载并准备数据集
+def formatting_prompts_func(example):
+    output_texts = []
+    for i in range(len(example['instruction'])):
+        text = f"### Instruction:\n{example['instruction'][i]}\n\n### Input:\n{example['input'][i]}\n\n### Response:\n{example['output'][i]}"
+        output_texts.append(text)
+    return output_texts
+
+dataset = load_dataset("json", data_files="data.json", split="train")
+
+# 3. 配置LoRA参数
+lora_config = LoraConfig(
+    r=8, # Rank
+    lora_alpha=32,
+    lora_dropout=0.1,
+    bias="none",
+    task_type="CAUSAL_LM",
+    target_modules=["q_proj", "k_proj", "v_proj", "o_proj"], # Llama-2 specific modules
+)
+
+# 4. 创建PEFT模型
+model = prepare_model_for_kbit_training(model)
+model = get_peft_model(model, lora_config)
+
+# 5. 配置训练参数
+output_dir = "./llama-2-7b-chat-json"
+training_args = TrainingArguments(
+    output_dir=output_dir,
+    per_device_train_batch_size=4,
+    gradient_accumulation_steps=4,
+    learning_rate=2e-4,
+    logging_steps=10,
+    max_steps=100, # for demo
+    save_strategy="epoch",
+    # num_train_epochs=1, # use max_steps for demo
+)
+
+# 6. 创建Trainer并开始训练
+trainer = SFTTrainer(
+    model=model,
+    train_dataset=dataset,
+    args=training_args,
+    peft_config=lora_config,
+    formatting_func=formatting_prompts_func,
+    max_seq_length=512,
+)
+
+trainer.train()
+
+# 7. 保存模型
+print("Saving LoRA adapter...")
+trainer.save_model(output_dir)
+print(f"LoRA adapter saved to {output_dir}")

requirements.txt

@@ -0,0 +1,7 @@
+torch
+transformers
+peft
+trl
+bitsandbytes
+datasets
+gradio