twin1shun/llm-jp-3-13b-finetune

実行手順

セットアップ

1. 必要なライブラリのインストール

!pip install -U bitsandbytes
!pip install -U transformers
!pip install -U accelerate
!pip install -U datasets
!pip install -U pef

from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
    BitsAndBytesConfig
)
from peft import PeftModel
import torch
from tqdm import tqdm
import re
import json
import gc

2. Hugging Faceのトークン取得

   以下はGoogle Colabでuserdataを使う例です（実行環境に合わせて適宜変更してください）。

from google.colab import userdata
HF_TOKEN = userdata.get('HF_TOKEN')

モデルの読み込み

model_id = "llm-jp/llm-jp-3-13b" 
adapter_id = "twin1shun/llm-jp-3-13b-finetune"

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4", # nf4は通常のINT4より精度が高く、ニューラルネットワークの分布に最適です
    bnb_4bit_compute_dtype=torch.bfloat16
)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    quantization_config=bnb_config,
    device_map="auto",
    token = HF_TOKEN
)

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

model = PeftModel.from_pretrained(model, adapter_id, token=HF_TOKEN)

データセットの読み込み

dataset = []
with open("./elyza-tasks-100-TV_0.jsonl", "r") as f:
  item = ""
  for line in f:
  line = line.strip()
  item += line
  if item.endswith("}"):
    datasets.append(json.loads(item))
    item = ""

モデルによるタスクの推論

results = []
for data in tqdm(datasets):

  input = data["input"]

  prompt = f"""### 指示
  {input}
  ### 回答
  """
    
  tokenized_input = tokenizer.encode(prompt, add_special_tokens=False, return_tensors="pt").to(model.device)
  attention_mask = torch.ones_like(tokenized_input)
  with torch.no_grad():
      outputs = model.generate(
          tokenized_input,
          attention_mask=attention_mask,
          max_new_tokens=100,
          do_sample=False,
          repetition_penalty=1.2,
          pad_token_id=tokenizer.eos_token_id
      )[0]
  output = tokenizer.decode(outputs[tokenized_input.size(1):], skip_special_tokens=True)

  results.append({"task_id": data["task_id"], "input": input, "output": output})

jsonlの生成

jsonl_id = re.sub(".*/", "", adapter_id)
with open(f"./{jsonl_id}-outputs.jsonl", 'w', encoding='utf-8') as f:
    for result in results:
        json.dump(result, f, ensure_ascii=False)  # ensure_ascii=False for handling non-ASCII characters
        f.write('\n')