mas_name stringclasses 7
values | llm_name stringclasses 5
values | benchmark_name stringclasses 8
values | trace_id int64 0 205 | trace dict | mast_annotation dict |
|---|---|---|---|---|---|
ChatDev | GPT-4o | ProgramDev | 0 | {"key":"ChatDev_ProgramDev_GPT4o","index":0,"trajectory":"[2025-31-03 19:09:41 INFO] **[Preprocessin(...TRUNCATED) | {"1.1":0,"1.2":0,"1.3":0,"1.4":0,"1.5":0,"2.1":0,"2.2":0,"2.3":0,"2.4":0,"2.5":0,"2.6":0,"3.1":0,"3.(...TRUNCATED) |
ChatDev | GPT-4o | ProgramDev | 1 | {"key":"ChatDev_ProgramDev_GPT4o","index":1,"trajectory":"[2025-31-03 19:30:18 INFO] **[Preprocessin(...TRUNCATED) | {"1.1":0,"1.2":0,"1.3":0,"1.4":0,"1.5":0,"2.1":0,"2.2":0,"2.3":0,"2.4":0,"2.5":0,"2.6":0,"3.1":0,"3.(...TRUNCATED) |
ChatDev | GPT-4o | ProgramDev | 2 | {"key":"ChatDev_ProgramDev_GPT4o","index":2,"trajectory":"[2025-31-03 19:48:48 INFO] **[Preprocessin(...TRUNCATED) | {"1.1":0,"1.2":0,"1.3":0,"1.4":0,"1.5":0,"2.1":0,"2.2":0,"2.3":0,"2.4":0,"2.5":0,"2.6":0,"3.1":0,"3.(...TRUNCATED) |
ChatDev | GPT-4o | ProgramDev | 3 | {"key":"ChatDev_ProgramDev_GPT4o","index":3,"trajectory":"[2025-31-03 20:04:23 INFO] **[Preprocessin(...TRUNCATED) | {"1.1":0,"1.2":0,"1.3":0,"1.4":0,"1.5":0,"2.1":0,"2.2":1,"2.3":1,"2.4":0,"2.5":0,"2.6":0,"3.1":0,"3.(...TRUNCATED) |
ChatDev | GPT-4o | ProgramDev | 4 | {"key":"ChatDev_ProgramDev_GPT4o","index":4,"trajectory":"[2025-31-03 20:15:44 INFO] **[Preprocessin(...TRUNCATED) | {"1.1":0,"1.2":0,"1.3":0,"1.4":0,"1.5":0,"2.1":0,"2.2":0,"2.3":0,"2.4":0,"2.5":0,"2.6":0,"3.1":0,"3.(...TRUNCATED) |
ChatDev | GPT-4o | ProgramDev | 5 | {"key":"ChatDev_ProgramDev_GPT4o","index":5,"trajectory":"[2025-31-03 20:40:53 INFO] **[Preprocessin(...TRUNCATED) | {"1.1":1,"1.2":0,"1.3":0,"1.4":0,"1.5":0,"2.1":0,"2.2":0,"2.3":0,"2.4":0,"2.5":0,"2.6":1,"3.1":0,"3.(...TRUNCATED) |
ChatDev | GPT-4o | ProgramDev | 6 | {"key":"ChatDev_ProgramDev_GPT4o","index":6,"trajectory":"[2025-31-03 20:46:17 INFO] **[Preprocessin(...TRUNCATED) | {"1.1":1,"1.2":0,"1.3":0,"1.4":0,"1.5":0,"2.1":0,"2.2":0,"2.3":0,"2.4":0,"2.5":0,"2.6":0,"3.1":0,"3.(...TRUNCATED) |
ChatDev | GPT-4o | ProgramDev | 7 | {"key":"ChatDev_ProgramDev_GPT4o","index":7,"trajectory":"[2025-31-03 20:55:23 INFO] **[Preprocessin(...TRUNCATED) | {"1.1":0,"1.2":0,"1.3":1,"1.4":0,"1.5":0,"2.1":0,"2.2":0,"2.3":0,"2.4":0,"2.5":0,"2.6":0,"3.1":0,"3.(...TRUNCATED) |
ChatDev | GPT-4o | ProgramDev | 8 | {"key":"ChatDev_ProgramDev_GPT4o","index":8,"trajectory":"[2025-31-03 21:05:25 INFO] **[Preprocessin(...TRUNCATED) | {"1.1":1,"1.2":0,"1.3":1,"1.4":0,"1.5":1,"2.1":0,"2.2":1,"2.3":1,"2.4":1,"2.5":1,"2.6":1,"3.1":1,"3.(...TRUNCATED) |
ChatDev | GPT-4o | ProgramDev | 9 | {"key":"ChatDev_ProgramDev_GPT4o","index":9,"trajectory":"[2025-31-03 21:19:38 INFO] **[Preprocessin(...TRUNCATED) | {"1.1":0,"1.2":0,"1.3":0,"1.4":0,"1.5":0,"2.1":0,"2.2":0,"2.3":0,"2.4":0,"2.5":0,"2.6":0,"3.1":0,"3.(...TRUNCATED) |
mcemri/MAST-Data(複製)
データの内容
マルチエージェントシステムの軌跡に、MAST の 14 の失敗モードの二値のアノテーションを LLM の判定で付けたデータ。1 行が 1 件の軌跡で、mas_name・llm_name・benchmark_name と、軌跡本体の trace、失敗モードごとの mast_annotation(1・0・null)が入る。人手のラベルは human_labelled の構成の 19 件(3 人のアノテーター)だけで、ラウンドごとに分類体系の版が違うため、コードはラウンドをまたいで比べられない。
データの概要
| 項目 | 値 |
|---|---|
| 行数 | 1,661(2 構成・2 分割の合計) |
| 列数 | 6 |
| 容量 | 202.24 MB(ファイル 4 本) |
| ファイルの形式 | json 2 |
| 公開 | public |
構成と分割
| 構成(config) | 分割(split)ごとの行数 | 列数 |
|---|---|---|
default |
train 1,642 |
6 |
human_labelled |
train 19 |
6 |
列(構成 default・分割 train。構成ごとに列が違う)
| 列 | 型 | 先頭の行の値 |
|---|---|---|
mas_name |
string |
ChatDev |
llm_name |
string |
GPT-4o |
benchmark_name |
string |
ProgramDev |
trace_id |
int64 |
0 |
trace |
struct{key: string, index: int64, trajectory: string} |
{"key":"ChatDev_ProgramDev_GPT4o","index":0,"tra… |
mast_annotation |
struct{1.1: int64, 1.2: int64, 1.3: int64, 1.4: int64, 1.5: int64, …} |
{"1.1":0,"1.2":0,"1.3":0,"1.4":0,"1.5":0,"2.1":0… |
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 にある。
複製について
mcemri/MAST-Data をリビジョン 95118ac951421753cf1deb87ddea3b01e693c41b 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/mcemri/MAST-Data
- 複製したリビジョン:
95118ac951421753cf1deb87ddea3b01e693c41b(2026-08-13 19:46:12+00:00) - 元のライセンス表記:
cc-by-4.0
論文
- 元のカードに論文へのリンクは見つからなかった
アノテーションの方法
(元のカードに該当の見出しが無いので、ラベルの付け方に触れた段落を抜き出した)
マルチエージェントシステム(MAS)の実行の軌跡に、Multi-Agent Systems Failure Taxonomy(MAST)でアノテーションを付けたもの。各レコードは、MAS、その背後の LLM、ベンチマークのタスク、軌跡の全体、そして 14 の MAST の失敗モードに対する二値のアノテーションを与える。
アノテーションは人手のラベル付けではなく、LLM の判定者(LLM judge)によって生成される。
MAD_human_labelled_dataset.json では、各ラウンドが分類体系の異なる改訂版を用いており(18、17、17、14 のモード)、コードはラウンドをまたいで比較できない。
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 (L86) | cc-by-4.0 | 確認済 | 生成元モデルの出力規約が別に及ぶ(Q4) / 原文注記: ラベルは LLM 判定で、人手の一致は 19 件でのみ検証され、ラウンドごとに分類の版が違う(18・17・17・14 モード。カード記載)。軌跡の長さは未測定。trace は dict で、入れ子が大きい |
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。
MAD: マルチエージェントシステムの軌跡データセット
マルチエージェントシステム(MAS)の実行の軌跡に、Multi-Agent Systems Failure Taxonomy(MAST)でアノテーションを付けたもの。各レコードは、MAS、その背後の LLM、ベンチマークのタスク、軌跡の全体、そして 14 の MAST の失敗モードに対する二値のアノテーションを与える。
コード: https://github.com/multi-agent-systems-failure-taxonomy/MAST
1642 件の軌跡 · 7 つの MAS フレームワーク · 8 つのベンチマーク · 5 つの LLM。
ファイル
| file | rows | |
|---|---|---|
MAD_full_dataset.json |
1642 | MAST のアノテーションを付けた軌跡 |
MAD_human_labelled_dataset.json |
19 | アノテーター間の一致の研究、3 人のアノテーター |
スキーマ
| field | type | |
|---|---|---|
mas_name |
string | AG2, MetaGPT, ChatDev, Magentic, AppWorld, HyperAgent, OpenManus |
llm_name |
string | GPT-4o, Claude, GPT-4o-mini, Qwen, CodeLlama |
benchmark_name |
string | ProgramDev, ProgramDev-v2, GSM, Olympiad, GAIA, MMLU, Test-C, SWE-Bench-Lite |
trace_id |
int | 設定の中での索引 |
trace |
dict | {key, index, trajectory} |
mast_annotation |
dict | 14 個のコード → 1、0、または、アノテーションが利用できない場合は null |
分類体系
仕様(Specification) — 1.1 タスク仕様への不服従(Disobey Task Specification) · 1.2 役割仕様への不服従(Disobey Role Specification) · 1.3 手順の反復(Step Repetition) · 1.4 会話履歴の喪失(Loss of Conversation History) · 1.5 終了条件の未認識(Unaware of Termination Conditions)
エージェント間の不整合(Inter-Agent Misalignment) — 2.1 会話のリセット(Conversation Reset) · 2.2 明確化の要求の欠如(Fail to Ask for Clarification) · 2.3 タスクからの逸脱(Task Derailment) · 2.4 情報の保留(Information Withholding) · 2.5 他のエージェントの入力の無視(Ignored Other Agent's Input) · 2.6 推論と行動の不一致(Reasoning-Action Mismatch)
タスク検証(Task Verification) — 3.1 早すぎる終了(Premature Termination) · 3.2 検証の欠如または不完全(No or Incomplete Verification) · 3.3 誤った検証(Incorrect Verification)
注記
アノテーションは人手のラベル付けではなく、LLM の判定者(LLM judge)によって生成される。
MAD_human_labelled_dataset.json では、各ラウンドが分類体系の異なる改訂版を用いており(18、17、17、14 のモード)、コードはラウンドをまたいで比較できない。
引用
@inproceedings{cemri2025mast,
title = {Why Do Multi-Agent LLM Systems Fail?},
author = {Cemri, Mert and Pan, Melissa Z. and Yang, Shuyi and Agrawal, Lakshya A. and Chopra, Bhavya and Tiwari, Rishabh and Keutzer, Kurt and Parameswaran, Aditya and Klein, Dan and Ramchandran, Kannan and Zaharia, Matei and Gonzalez, Joseph E. and Stoica, Ion},
booktitle = {Advances in Neural Information Processing Systems},
year = {2025}
}
原文(英語)
MAD: Multi-Agent System Traces Dataset
Execution traces from multi-agent systems (MAS), annotated with the Multi-Agent Systems Failure Taxonomy (MAST). Each record gives the MAS, the LLM behind it, the benchmark task, the full trace, and binary annotations for the 14 MAST failure modes.
Code: https://github.com/multi-agent-systems-failure-taxonomy/MAST
1642 traces · 7 MAS frameworks · 8 benchmarks · 5 LLMs.
Files
| file | rows | |
|---|---|---|
MAD_full_dataset.json |
1642 | traces with MAST annotations |
MAD_human_labelled_dataset.json |
19 | inter-annotator agreement study, 3 annotators |
Schema
| field | type | |
|---|---|---|
mas_name |
string | AG2, MetaGPT, ChatDev, Magentic, AppWorld, HyperAgent, OpenManus |
llm_name |
string | GPT-4o, Claude, GPT-4o-mini, Qwen, CodeLlama |
benchmark_name |
string | ProgramDev, ProgramDev-v2, GSM, Olympiad, GAIA, MMLU, Test-C, SWE-Bench-Lite |
trace_id |
int | index within its config |
trace |
dict | {key, index, trajectory} |
mast_annotation |
dict | 14 codes → 1, 0, or null where the annotation is unavailable |
Taxonomy
Specification — 1.1 Disobey Task Specification · 1.2 Disobey Role Specification ·
1.3 Step Repetition · 1.4 Loss of Conversation History · 1.5 Unaware of Termination Conditions
Inter-Agent Misalignment — 2.1 Conversation Reset · 2.2 Fail to Ask for Clarification ·
2.3 Task Derailment · 2.4 Information Withholding · 2.5 Ignored Other Agent's Input ·
2.6 Reasoning-Action Mismatch
Task Verification — 3.1 Premature Termination · 3.2 No or Incomplete Verification ·
3.3 Incorrect Verification
Notes
Annotations are produced by an LLM judge, not by human labelling.
In MAD_human_labelled_dataset.json, each round uses a different revision of the taxonomy
(18, 17, 17 and 14 modes) and the codes are not comparable across rounds.
Citation
@inproceedings{cemri2025mast,
title = {Why Do Multi-Agent LLM Systems Fail?},
author = {Cemri, Mert and Pan, Melissa Z. and Yang, Shuyi and Agrawal, Lakshya A. and Chopra, Bhavya and Tiwari, Rishabh and Keutzer, Kurt and Parameswaran, Aditya and Klein, Dan and Ramchandran, Kannan and Zaharia, Matei and Gonzalez, Joseph E. and Stoica, Ion},
booktitle = {Advances in Neural Information Processing Systems},
year = {2025}
}
- Downloads last month
- 42