HYUKI5's picture
LulaCola/AgentProcessBench@cd81f326aece を複製し、カードを日本語で書く
d790510 verified
|
Raw History Blame Contribute Delete
6.36 kB
---
license: mit
pretty_name: LulaCola/AgentProcessBench (mirror)
source_datasets:
- LulaCola/AgentProcessBench
tags:
- laya
- laya-finetune
- mirror
- domain-general
- use-conditional
- genre-dialogue
- label-multiclass
- annot-unknown
task_categories:
- question-answering
- text-generation
task_ids:
- extractive-qa
- open-domain-qa
- task-planning
configs:
- config_name: bfcl
data_files:
- split: test
path: bfcl/test.jsonl
- config_name: gaia_dev
data_files:
- split: test
path: gaia_dev/test.jsonl
- config_name: hotpotqa
data_files:
- split: test
path: hotpotqa/test.jsonl
- config_name: tau2
data_files:
- split: test
path: tau2/test.jsonl
---
# LulaCola/AgentProcessBench(複製)
## データの内容
ツールを使うエージェントの、プロセスレベル(ステップごと)の評価のための、エージェントの軌跡のデータ。1 件が 1 つの完全な軌跡で、`messages`・`tools`・`question` に会話と定義が、`step_labels` にステップごとのラベル、`final_label` に軌跡全体のラベルが入る。bfcl・gaia_dev・hotpotqa・tau2 の 4 つの構成に各 250 件ずつあり、ラベルの三値の意味はカードに書かれていない。論文は arXiv:2603.14465。
## データの概要
| 項目 | 値 |
|---|---|
| 行数 | 1,000(4 構成・4 分割の合計) |
| 列数 | 12〜13(構成による) |
| 容量 | 40.11 MB(ファイル 6 本) |
| ファイルの形式 | jsonl 4 |
| 公開 | public |
### 構成と分割
| 構成(config) | 分割(split)ごとの行数 | 列数 |
|---|---|---:|
| `bfcl` | `test` 250 | 13 |
| `gaia_dev` | `test` 250 | 12 |
| `hotpotqa` | `test` 250 | 12 |
| `tau2` | `test` 250 | 13 |
### 列(構成 `bfcl`・分割 `test`。構成ごとに列が違う)
| 列 | 型 | 先頭の行の値 |
|---|---|---|
| `total_index` | `int64` | `500` |
| `query_index` | `int64` | `0` |
| `sample_index` | `int64` | `0` |
| `question` | `string` | `Hey, can you set up a new directory named 'Proje…` |
| `data_source` | `string` | `bfcl_multi_turn_base` |
| `ground_truth` | `list<list<string>>` | `[5 件]` |
| `answer_text` | `string` | `Great! The 'Projects' directory has been success…` |
| `messages` | `list<json>` | `[55 件]` |
| `tool_metrics` | `json` | `{pwd, mkdir, ls, cd, …}` |
| `tools` | `list<struct{name: string, description: string, parameters: struct{…}, response: struct{…}}>` | `[18 件]` |
| `step_labels` | `json` | `{3, 5, 7, 9, …}` |
| `final_label` | `int64` | `-1` |
| `task_description` | `null` | `null` |
行数と列は HF のデータセットビューアの集計(元のデータセットの現在の版で数えたもの)。全データセットの一覧は [HYUKI5/dataset-catalog](https://huggingface.co/datasets/HYUKI5/dataset-catalog) にある。
## 複製について
[`LulaCola/AgentProcessBench`](https://huggingface.co/datasets/LulaCola/AgentProcessBench) をリビジョン `cd81f326aece8b0a3f6359e20df370690d3a98bf` の時点でそのまま複製したもの。中身は変えていない。判断モデル Laya(`convaiinnovations/laya`)の微調整に使う候補として、調査メモ(`Laya 学習用データセット調査`)に載せたデータを集めている。
## 出典
- 元のデータセット: https://huggingface.co/datasets/LulaCola/AgentProcessBench
- 複製したリビジョン: `cd81f326aece8b0a3f6359e20df370690d3a98bf`(2026-03-18 01:40:42+00:00)
- 元のライセンス表記: `mit`
## 論文
- https://arxiv.org/abs/2603.14465
## アノテーションの方法
元のカードにアノテーションの節が無かった。下の「調査メモの記述」と「元のカード」を参照。
## 微調整に使うときの注意
- 調査メモの台帳での区分: B(継承条件・gated・モデル由来の条件・上流の規約などが付く)。区分は記載の文言で機械的に付けたもので、法的な判定ではない。学習に入れる前に元のカードとライセンスの原文で確かめる。
- Laya の入力は 2,048 token まで(ModernBERT-large / mmBERT)。長い本文は先頭から黙って切られるので、切れる行は事前に除くか、要約・切り出しをしてから入れる。
- 分割は記事・文書単位で行い、同じ本文から作った別の問いを学習側と較正側に分けない。温度と閾値は学習に使っていない分割で別々に当てる。
## 調査メモの記述
| メモ | 節 | ライセンス(記載) | 確認 | 注記 |
|---|---|---|---|---|
| `agent_trajectories_thought_logs.md` | Q3. エージェントの失敗の種類・責任ステップ・ステップ単位の評価が付いたデータは何があるか (L90) | mit | 確認済 | 生成元モデルの出力規約が別に及ぶ(Q4) / 原文注記: 1,000 件と小さいが、ステップ単位なので学習事例は 8,509 件。どのモデルが軌跡を生成したかは abs ページで確認できず。三値の意味は未確認 |
## 元のカード(日本語訳)
以下は元のデータセットのカードの本文を日本語に訳したもの(機械翻訳)。訳に疑問があるときは、末尾の原文を確かめる。
---
# AgentProcessBench
AgentProcessBench は、ツールを使うエージェントのプロセスレベルの評価のためのベンチマークである。各例は、複数ターンのメッセージ、ツールの定義、ツール使用の軌跡、参照出力、ステップごとのプロセスラベルを含む、エージェントの完全な軌跡である。
ベンチマークには合計 1,000 件の軌跡が含まれ、各サブセットから 250 件ずつ含まれる。
- `bfcl`
- `gaia_dev`
- `hotpotqa`
- `tau2`
arxiv.org/abs/2603.14465
<details>
<summary>原文(英語)</summary>
# AgentProcessBench
AgentProcessBench is a benchmark for process-level evaluation of tool-using agents. Each example is a full agent trajectory with multi-turn messages, tool definitions, tool-use traces, reference outputs, and step-wise process labels.
The benchmark contains 1,000 trajectories in total, with 250 examples from each subset:
- `bfcl`
- `gaia_dev`
- `hotpotqa`
- `tau2`
arxiv.org/abs/2603.14465
</details>