大模型基于检索增强的微调-RAFT

import os
import datasets
import pandas as pd
from llama_index.packs.raft_dataset import RAFTDatasetPack
from llama_index.llms.ollama import Ollama
from llama_index.embeddings.ollama import OllamaEmbedding

def gen_qa_raft_dataset():
    llm = Ollama(model="qwen3:32b", request_timeout=60.0)

    embed_model = OllamaEmbedding(model_name="modelscope.cn/yingda/gte-Qwen2-1.5B-instruct-GGUF:latest")

    sub_dir = 'data/laws/'

    df_list = []
    for file_path in os.listdir(sub_dir):
        print(file_path)
        file_path = os.path.join(sub_dir, file_path)
        raft_dataset = RAFTDatasetPack(file_path,
                                       llm=llm,
                                       embed_model=embed_model,
                                       num_questions_per_chunk=3,
                                       num_distract_docs=3,
                                       chunk_size=2048)
        dataset = raft_dataset.run()
        df = dataset.to_pandas()
        df['messages'] = df.apply(lambda row: [{'content': row['instruction'], 'role': 'user'},
                                               {'content': row['cot_answer'], 'role': 'assistant'}], axis=1)
        df_list.append(df)

    df_all = pd.concat(df_list)
    dataset = datasets.Dataset.from_pandas(df_all)

    dataset.to_csv("data/法律文书_raft_dataset.csv", encoding='utf-8')

tokenizer.chat_template = DEFAULT_CHAT_TEMPLATE
SYSTEM_PROMPT = """你是一位很有能力的问题解答者，能够根据问题和相关背景提供答案。"""
def apply_chat_template(example, tokenizer):
    messages = example["messages"]
    messages = messages.replace('\n',',')
    messages = eval(messages)
    # We add an empty system message if there is none
    if messages[0]["role"] != "assistant":
        messages.insert(0, {"role": "system", "content": SYSTEM_PROMPT})
    example["text"] = tokenizer.apply_chat_template(messages, tokenize=False)

    return example
column_names = list(dataset.features)
raw_datasets = dataset.map(apply_chat_template,
                                num_proc=cpu_count(),
                                fn_kwargs={"tokenizer": tokenizer},
                                remove_columns=column_names,
                                desc="Applying chat template",)
raw_datasets = raw_datasets.train_test_split(test_size=0.1)
# create the splits
train_dataset = raw_datasets["train"]
eval_dataset = raw_datasets["test"]

# 加载模型 - 使用Unsloth的FastLanguageModel加载预训练的Qwen3-8B模型
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name='Qwen/Qwen3-0.6B',  # 模型路径
    dtype=torch.float16,  # 使用float16数据类型以减少内存占用
    max_seq_length=2048,  # 设置最大序列长度为2048
    load_in_4bit=False  # 使用4位量化加载模型，进一步减少显存占用
)
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,
    train_dataset = train_dataset,
    eval_dataset = eval_dataset,
    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.
        max_steps = 30,
        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
    ),
)
tokenizer.pad_token = tokenizer.eos_token
train_result = trainer.train()