1.unsloth-Burden-based.py

from unsloth import FastLanguageModel
import os
import torch
from datasets import load_dataset
from trl import SFTConfig, SFTTrainer
import random # 【魔改】导入 random 库用于乱序

# os.environ["UNSLOTH_VLLM_STANDBY"] = "1"

# --- 本地路径配置 (无需更改) ---
# my_load_model = "Qwen3-30B-A3B-Thinking-2507"
my_load_model = "Qwen3-4B-Thinking-2507"
my_model_name = "QiMing-Polaris"
max_seq_length = 4096

print(f"Dataset: {my_model_name}")

local_model_path = f"/home/aifeifei/AI_Data/develop/mini_tang/modules/{my_load_model}"
local_data_dir = f"{my_model_name}"
local_data_file = os.path.join(local_data_dir, f"{my_model_name}.jsonl")
final_model_path = f"{my_model_name}-{my_load_model}_burden_trained_lora" # 【魔改】改个名，标记这是负重训练版

# --- 配置结束 ---


# 1. 加载模型和分词器 (无需更改)

dtype = (
    None
)
load_in_4bit = True

print(f"✅ 步骤 1/6: 正在从本地路径 '{local_model_path}' 加载模型和分词器...")
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name=local_model_path,
    max_seq_length=max_seq_length,
    dtype=dtype,
    load_in_4bit=load_in_4bit,
    full_finetuning=False,
)
print("🎉 模型加载完成！")

# 2. 配置 LoRA (无需更改)
print("✅ 步骤 2/6: 正在配置 LoRA 适配器...")
model = FastLanguageModel.get_peft_model(
    model,
    r=8,
    target_modules=[
        "q_proj", "k_proj", "v_proj", "o_proj",
        "gate_proj", "up_proj", "down_proj",
    ],
    lora_alpha=16,
    lora_dropout=0,
    bias="none",
    use_gradient_checkpointing="unsloth",
    random_state=3407,
    use_rslora=False,
    loftq_config=None,
)
print("🎉 LoRA 配置完成！")

# 3. 加载和准备数据集 (【魔改】核心部分)
alpaca_prompt = """Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.

### Instruction:
{}

### Input:
{}

### Response:
{}"""


EOS_TOKEN = tokenizer.eos_token

# =================================================================================
# 【魔改】 注入“负重训练”逻辑！
# =================================================================================
def apply_burden(text, burden_ratio=0.7):
    """
    给一段文本绑上“铅袋”：按一定比例打乱词序。
    """
    words = text.split(' ')
    # 只有当单词数量大于3时才进行乱序，避免太短的文本失去意义
    if len(words) > 3:
        num_to_shuffle = int(len(words) * burden_ratio)
        # 随机选择一些单词的索引
        indices_to_shuffle = random.sample(range(len(words)), num_to_shuffle)
        # 只打乱这些被选中的单词
        shuffled_subset = [words[i] for i in indices_to_shuffle]
        random.shuffle(shuffled_subset)
        # 把打乱后的单词放回原位
        shuffled_words = list(words) # 创建一个副本
        for i, original_index in enumerate(indices_to_shuffle):
            shuffled_words[original_index] = shuffled_subset[i]
        return ' '.join(shuffled_words)
    return text

def formatting_prompts_func(examples):
    all_texts = []
    for i in range(len(examples["instruction"])):
        instruction = examples["instruction"][i]
        input_text = examples["input"][i]
        # 【魔改】 output 保持原样，是我们的“完美答案”
        output_text = examples["output"][i]

        # 【魔改】给 instruction 和 input 绑上“铅袋”！
        burdened_instruction = apply_burden(instruction)
        burdened_input = apply_burden(input_text)

        # 【魔改】用“七零八落”的输入，去训练模型得到“规整”的输出
        text = alpaca_prompt.format(burdened_instruction, burdened_input, output_text) + EOS_TOKEN
        all_texts.append(text)
    return {"text": all_texts}
# =================================================================================


print(f"✅ 步骤 3/6: 正在从HF '{local_data_file}' 加载并应用“负重训练”处理...")
dataset = load_dataset("json", data_files=local_data_file, split="train")
dataset = dataset.map(
    formatting_prompts_func,
    batched=True,
    remove_columns=dataset.column_names,
    load_from_cache_file=False,
)
print(f"🎉 数据集处理完成！总共生成了 {len(dataset)} 条“负重”训练样本。")
print("来看一个“绑了铅袋”的样本长啥样：\n")
print(dataset[0]['text']) # 打印第一条看看效果

# 4. 配置训练参数并开始训练
print("\n✅ 步骤 4/5: 开始模型微调...")

trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    train_dataset=dataset,
    dataset_text_field="text",
    max_seq_length=max_seq_length,
    dataset_num_proc=8,
    packing=False,
    args=SFTConfig(
        per_device_train_batch_size=4,
        gradient_accumulation_steps=2,
        dataloader_num_workers=4,
        dataloader_pin_memory=True,
        warmup_steps=25,
        num_train_epochs=10,
        learning_rate=2e-5,
        fp16=not torch.cuda.is_bf16_supported(),
        bf16=torch.cuda.is_bf16_supported(),
        logging_steps=5,
        optim="paged_adamw_8bit",
        lr_scheduler_type="cosine",
        warmup_ratio=0.03,
        weight_decay=0.01,
        seed=3407,
        output_dir = f"output/{final_model_path}", # 建议加上输出目录，方便续训
        report_to="none",
    ),
)
trainer.train()

# 5. 保存并测试 (无需更改)
print("\n✅ 步骤 5/5: 微调完成...")
model.save_pretrained(final_model_path)
tokenizer.save_pretrained(final_model_path)
print(f"🎉 “负重训练”版 LoRA 模型已保存到 '{final_model_path}' 文件夹。")