00201 Hugging Face Llama Recipes - Fine Tuning

前言

简单的介绍如何Fine Tuning Llama。

src link: https://github.com/LuYF-Lemon-love/fork-huggingface-llama-recipes

Llama Models: https://modelscope.cn/organization/LLM-Research

Operating System: Ubuntu 22.04.4 LTS

参考文档

1. NLP Course - Fine Tuning

介绍

通常只在模型上运行推理是不够的。很多时候，您需要在一些自定义数据集上对模型进行微调。以下是一些脚本，展示了如何对模型进行微调。

peft finetuning

src link: https://github.com/LuYF-Lemon-love/fork-huggingface-llama-recipes/blob/main/fine_tune/peft_finetuning.py

# This example is a very quick showcase of partial fine-tuning the Llama 3.2 3B model
# on the IMDB dataset using QLoRA with bitsandbytes.

# In order to run this example, you'll need to install peft, trl, and bitsandbytes:
# pip install peft trl bitsandbytes

import torch
from datasets import load_dataset

from trl import SFTConfig, SFTTrainer
from peft import LoraConfig
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig

from modelscope import snapshot_download

model_dir = snapshot_download('LLM-Research/Llama-3.2-3B')

tokenizer = AutoTokenizer.from_pretrained(model_dir)
model = AutoModelForCausalLM.from_pretrained(model_dir)

tokenizer.pad_token = tokenizer.eos_token

dataset = load_dataset("imdb", split="train")

sft_config = SFTConfig(
    dataset_text_field="text",
    per_device_train_batch_size=4,
    max_seq_length=20,
    num_train_epochs=3,
    output_dir="./results",
    logging_dir='./logs',
    logging_steps=10,
)

QLoRA = True
if QLoRA:
    quantization_config = BitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_compute_dtype=torch.float16,
        bnb_4bit_quant_type="nf4"
    )
    
    lora_config = LoraConfig(
        r=8,
        target_modules="all-linear",
        bias="none",
        task_type="CAUSAL_LM",
    )
else:
    lora_config = None

trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    args=sft_config,
    peft_config=lora_config,
    train_dataset=dataset,
)

trainer.train()

结语

第二百零一篇博文写完，开心！！！！

今天，也是充满希望的一天。