metadata
annotations_creators:
- synthetic
language:
- en
license:
- agpl-3.0
v-Fable: Synthetic Fable-5 Traces Dataset
📊 データセット概要
このデータセットは Glint-Research/Fable-5-traces の統計分布を完全再現した合成データです。 AIエージェント(claude-fable-5)のコード生成・ツール使用のトレースを、元データと同一の分布で生成しています。
| 項目 | 値 |
|---|---|
| レコード数 | 100,000 |
| ファイル | v_fable.jsonl |
| セッション数 | 60 |
| モデル | claude-fable-5 |
| フォーマット | JSONL (10カラム) |
| ライセンス | AGPL-3.0 |
📋 カラム構成
| カラム | 説明 |
|---|---|
uid |
ユニークID({session_uuid}#{sequence}) |
source_file |
元のソースファイルパス |
session |
セッションUUID |
model |
モデル名(claude-fable-5) |
context |
会話コンテキスト(最大7,022文字) |
cot |
チェーン・オブ・ソート(思考過程) |
output_type |
text または tool_use |
output |
実際の出力(メッセージ or ツール呼び出し) |
completion |
<think>タグ付き完全な応答 |
origin |
生成元(local / hf) |
📈 統計情報
| 指標 | 元データ | 合成データ |
|---|---|---|
| レコード数 | 4,665 | 100,000 |
| % text | 18.6% | 18.7% |
| % tool_use | 81.4% | 81.3% |
| CoT 平均長 | 2,669文字 | 2,634文字 |
| CoT 中央値 | 2,365文字 | 2,633文字 |
| completion 平均長 | 3,754文字 | 2,945文字 |
| Bash | 1,544回 | 1,543回 |
| Edit | 960回 | 942回 |
| Read | 443回 | 465回 |
| Write | 311回 | 311回 |
🚀 使い方
from datasets import load_dataset
# データセットの読み込み
ds = load_dataset("summerMC/v-Fable", split="train")
print(len(ds)) # 4665
print(ds[0]["output_type"]) # "text" or "tool_use"
print(ds[0]["cot"][:200]) # Chain of thought
# ファインチューニング用にフォーマット
train_data = []
for r in ds:
train_data.append({
"input": r["context"],
"output": r["completion"],
})
🔧 生成方法
合成データは fable5_generator.py を使って生成されています:
pip install numpy
python3 fable5_generator.py --sessions 60 --records 4665 --output v_fable.jsonl --seed 42
📝 ライセンス
AGPL-3.0
🙏 クレジット
- 元データセット: Glint-Research/Fable-5-traces
- 統計解析・合成生成: v-Fable pipeline