看不懂代码如何微调DeepSeek-R1蒸馏的Llama-8B模型？

%%capture# Normally using pip install unsloth is enough
# Temporarily as of Jan 31st 2025, Colab has some issues with Pytorch# Using pip install unsloth will take 3 minutes, whilst the below takes <1 minute:!pip install --no-deps bitsandbytes accelerate xformers==0.0.29 peft trl triton!pip install --no-deps cut_cross_entropy unsloth_zoo!pip install sentencepiece protobuf datasets huggingface_hub hf_transfer!pip install --no-deps unsloth

from unsloth import FastLanguageModelimport torch
model, tokenizer = FastLanguageModel.from_pretrained(    model_name = "unsloth/DeepSeek-R1-Distill-Llama-8B-unsloth-bnb-4bit",    max_seq_length = 512,  # 减小到512    dtype = None,    load_in_4bit = True,)

model = FastLanguageModel.get_peft_model(    model,    r = 8,   # 减小LoRA秩    target_modules = ["q_proj", "k_proj", "v_proj"],  # 只保留核心模块    lora_alpha = 8,    lora_dropout = 0,    bias = "none",    use_gradient_checkpointing = "unsloth",    random_state = 3407,)

from datasets import load_dataset
def formatting_prompts_func(examples):    texts = []    for example in examples["text"]:        text = example + EOS_TOKEN        texts.append(text)    return {"text": texts}
# 加载数据集并划分dataset = load_dataset("text", data_files={"train": "jud.txt"})# 划分训练集和验证集,验证集占比0.1(可以根据需要调整)split_dataset = dataset["train"].train_test_split(test_size=0.1, seed=42)dataset = {    "train": split_dataset["train"],    "validation": split_dataset["test"]  # test部分用作validation}

from transformers import AutoTokenizer
# tokenizer部分保持不变tokenizer = AutoTokenizer.from_pretrained("unsloth/DeepSeek-R1-Distill-Llama-8B-unsloth-bnb-4bit")
# 加载txt数据集dataset = load_dataset(    "text",    data_files={"train": "jud.txt"},    split="train")
# 划分数据集split_dataset = dataset.train_test_split(test_size=0.1, seed=42)train_dataset = split_dataset["train"]val_dataset = split_dataset["test"]
# 查看数据结构print("数据集列名:", train_dataset.column_names)print("\n第一条数据样例:", train_dataset[0])
def format_text(examples):    """处理纯文本数据"""    formatted_texts = []    for text in examples["text"]:        # 为每个文本片段添加EOS token        formatted_text = text + tokenizer.eos_token        formatted_texts.append(formatted_text)    return {"text": formatted_texts}
# 应用数据处理train_dataset = train_dataset.map(    format_text,    batched=True,    remove_columns=train_dataset.column_names)
val_dataset = val_dataset.map(    format_text,    batched=True,    remove_columns=val_dataset.column_names)
# 验证输出print("\n处理后的训练集样例:")print(train_dataset[0]["text"][:200], "...") # 只打印前200个字符print("\n处理后的验证集样例:")print(val_dataset[0]["text"][:200], "...")

from trl import SFTTrainerfrom transformers import TrainingArgumentsfrom unsloth import is_bfloat16_supported
trainer = SFTTrainer(    model = model,    tokenizer = tokenizer,    train_dataset = train_dataset,    eval_dataset = val_dataset,    dataset_text_field = "text",    max_seq_length = 512,  # 减小序列长度    dataset_num_proc = 1,    packing = True,    args = TrainingArguments(        per_device_train_batch_size = 4,  # 增加batch size        gradient_accumulation_steps = 2,   # 减少梯度累积        warmup_steps = 5,        num_train_epochs = 1,    # 只训练1轮        learning_rate = 3e-4,    # 稍微提高学习率        fp16 = not is_bfloat16_supported(),        bf16 = is_bfloat16_supported(),        logging_steps = 1,        optim = "adamw_8bit",        weight_decay = 0.01,        lr_scheduler_type = "linear",        seed = 3407,        output_dir = "outputs",        report_to = "none",    ),)