README.md

metadata

license: apache-2.0
datasets:
  - Dongwookss/q_a_korean_futsal
language:
  - ko
tags:
  - unsloth
  - trl
  - transformer

Model Name : 풋풋이(futfut)

Model Concept

• 풋살 도메인 친절한 도우미 챗봇을 구축하기 위해 LLM 파인튜닝과 RAG를 이용하였습니다.
• Base Model : zephyr-7b-beta
• 풋풋이의 말투는 '해요'체를 사용하여 말끝에 '얼마든지 물어보세요! 풋풋!'로 종료합니다.

Summary:

• Unsloth 패키지를 사용하여 LoRA 진행하였습니다.

• SFT Trainer를 통해 훈련을 진행

• 활용 데이터

  • q_a_korean_futsal
    • 말투 학습을 위해 '해요'체로 변환하고 인삿말을 넣어 모델 컨셉을 유지하였습니다.

• Environment : Colab 환경에서 진행하였으며 L4 GPU를 사용하였습니다.

How to use

Model Load

#!pip install transformers==4.40.0 accelerate
import os
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = 'Dongwookss/small_fut_final'

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model.eval()

Query

from transformers import TextStreamer
PROMPT = '''Below is an instruction that describes a task. Write a response that appropriately completes the reques문"

messages = [
    {"role": "system", "content": f"{PROMPT}"},
    {"role": "user", "content": f"{instruction}"}
    ]

input_ids = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

terminators = [
    tokenizer.eos_token_id,
    tokenizer.convert_tokens_to_ids("<|eot_id|>")
]

text_streamer = TextStreamer(tokenizer)
_ = model.generate(
    input_ids,
    max_new_tokens=4096,
    eos_token_id=terminators,
    do_sample=True,
    streamer = text_streamer,
    temperature=0.6,
    top_p=0.9,
    repetition_penalty = 1.1
)

Fine-Tuning with Unsloth(SFT Trainer)

from unsloth import FastLanguageModel
import torch
from trl import SFTTrainer
from transformers import TrainingArguments

max_seq_length = 256
dtype = None
load_in_4bit = False
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="HuggingFaceH4/zephyr-7b-beta",
    max_seq_length=max_seq_length,
    dtype=dtype,
    load_in_4bit=load_in_4bit,
    #token = ,
)

model = FastLanguageModel.get_peft_model(
    model,
    r=32,
    lora_alpha=64,
    lora_dropout=0.05, 
    target_modules=[
        "q_proj",
        "k_proj",
        "v_proj",
        "o_proj",
        "gate_proj",
        "up_proj",
        "down_proj",
    ],  # 타겟 모듈
    bias="none",
    use_gradient_checkpointing="unsloth",
    random_state=123,
    use_rslora=False,
    loftq_config=None,
)

tokenizer.padding_side = "right"

trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    train_dataset=dataset,
    dataset_text_field="text",
    max_seq_length=max_seq_length,
    dataset_num_proc=2,
    packing=False,
    args=TrainingArguments(
        per_device_train_batch_size=20,
        gradient_accumulation_steps=2,
        warmup_steps=5,
        num_train_epochs=3,
        max_steps = 1761,
        logging_steps = 10,
        learning_rate=2e-5,
        fp16=not torch.cuda.is_bf16_supported(),
        bf16=torch.cuda.is_bf16_supported(),
        optim="adamw_8bit",
        weight_decay=0.01,
        lr_scheduler_type="cosine",
        seed=123,
        output_dir="outputs",
    ),
)

trainer.train()