README.md

---
license: apache-2.0
datasets:
- kinokokoro/ichikara-instruction-003
data_license:
- cc
language:
- ja
base_model:
- llm-jp/llm-jp-3-13b
library_name: transformers
tags:
- text-generation-inference
- transformers
---

# Sample Use
```python
from tqdm import tqdm
import json
import os

MODEL_DIR = os.path.join(BASE_DIR, "fine_tuned_model")


def generate_predictions(model, tokenizer, input_file, output_file):
    # バッチ処理の追加
    BATCH_SIZE = 16  # バッチサイズの設定

    print(f"入力ファイルを読み込み中: {input_file}")
    tasks = []
    with open(input_file, 'r', encoding='utf-8') as f:
        for line in f:
            tasks.append(json.loads(line))

    results = []
    print("推論を実行中...")

    # バッチ処理
    for i in tqdm(range(0, len(tasks), BATCH_SIZE)):
        batch_tasks = tasks[i:i + BATCH_SIZE]
        prompts = [f"入力: {task['input']}\n出力: " for task in batch_tasks]

        # バッチでの推論
        inputs = tokenizer(
            prompts,
            return_tensors="pt",
            padding=True,
            truncation=True,
            max_length=512
        )

        with torch.no_grad():
            outputs = model.generate(
                inputs.input_ids,
                max_length=512,
                temperature=0.9,
                do_sample=False,
                repetition_penalty=1.2,
                pad_token_id=tokenizer.pad_token_id,
                top_k=50,
                top_p=0.95,
                early_stopping=True,    # 早期停止を有効化
                use_cache=True          # キャッシュを使用
            )

        # バッチ出力の処理
        for k, task in enumerate(batch_tasks):  # 各タスクについてループ
          output_index = k  # インデックスはタスクごとに1つだけ
          if output_index < len(outputs):  # 範囲外アクセスを防ぐ
              generated_text = tokenizer.decode(outputs[output_index], skip_special_tokens=True)
              output_text = generated_text.split("出力: ")[-1].strip()
              results.append({
                  "task_id": task["task_id"],  # 正しいタスクIDを取得
                  "output": output_text       # 対応する出力
              })

    print(f"結果を保存中: {output_file}")
    with open(output_file, 'w', encoding='utf-8') as f:
        for result in results:
            json.dump(result, f, ensure_ascii=False)
            f.write('\n')

def main():
    # GPUメモリのクリア
    torch.cuda.empty_cache()

    # 時間計測の追加
    import time
    start_time = time.time()

    model, tokenizer = load_model()
    input_file = "{$file_path}"
    output_file = os.path.join(BASE_DIR, "{$file_path}")

    generate_predictions(model, tokenizer, input_file, output_file)

    # 実行時間の表示
    elapsed_time = time.time() - start_time
    print(f"総実行時間: {elapsed_time / 60:.2f}分")

if __name__ == "__main__":
    main()
```