r-yuba62/llm-jp-3-13b-finetune

Summary

This model was created as part of the creation of a submitted model for the competition of [the Matsuo Lab Large Scale Language Modeling Course 2024] (https://weblab.t.u-tokyo.ac.jp/lecture/course-list/large-language-model/) .

Uploaded model

Developed by: r-yuba License: apache-2.0 Finetuned from model : llm-jp/llm-jp-3-13b

How to Usage

To Install Packages

!pip install -U transformers
!pip install -U bitsandbytes
!pip install -U accelerate
!pip install -U datasets
!pip install -U peft
!pip install -U trl==0.12.0

Methods of Inference

from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
    BitsAndBytesConfig,
)
from peft import PeftModel
import torch

HF_TOKEN = "AVAILABLE YOUR-HF-TOKEN"

model_name = "llm-jp/llm-jp-3-13b"
adapter_name = "r-yuba62/llm-jp-3-13b-finetune"

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
)

# モデルとトークナイザーをロード
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    quantization_config=bnb_config,
    device_map="auto",
    token=HF_TOKEN
)

tokenizer = AutoTokenizer.from_pretrained(
    model_name, 
    trust_remote_code=True, 
    token=HF_TOKEN
)

# PEFTアダプターを適用
model = PeftModel.from_pretrained(model, adapter_name, token=HF_TOKEN)

def get_response(input_text):
    prompt = f"""### 指示
    {input_text}
    ### 回答：
    """
    # トークナイズ処理
    tokenized_input = tokenizer(
        prompt,
        return_tensors="pt",
        padding=True,
        truncation=True,
        max_length=512
    )

    input_ids = tokenized_input["input_ids"].to(model.device)
    attention_mask = tokenized_input["attention_mask"].to(model.device)

    # モデル生成
    with torch.no_grad():
        outputs = model.generate(
            input_ids=input_ids,
            attention_mask=attention_mask,
            max_new_tokens=512,
            do_sample=False,
            repetition_penalty=1.2,
            pad_token_id=tokenizer.eos_token_id
        )

    # 出力のデコード
    output_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return output_text

input_text = "xxxを教えてください"
response = get_response(input_text)
print(response)

Base Model

base_model:- llm-jp/llm-jp-3-13b

Instruction tuning

The models have been fine-tuned on the following datasets.

Language	Dataset	description
Japanese	ichikara-instruction-003-001-1.json	A manually constructed instruction dataset

データセット作成チーム： 関根聡, 安藤まや, 後藤美知子, 鈴木久美, 河原大輔, 井之上直也, 乾健太郎. ichikara-instruction: LLMのための日本語インストラクションデータの構築. 言語処理学会第30回年次大会(2024)

License

Apache License, Version 2.0