Dataset Viewer
Auto-converted to Parquet Duplicate
trajectory_id
stringlengths
26
51
task_type
stringclasses
3 values
llm_model
stringclasses
3 values
critical_failure_step
int64
1
30
critical_failure_module
stringclasses
6 values
failure_types
listlengths
1
1
failure_reasonings
listlengths
1
1
failure_modules
listlengths
1
1
num_steps
int64
2
60
trajectory_length
int64
5k
6.21M
full_trajectory
stringlengths
5.19k
6.21M
step_annotations
stringlengths
60
1.33k
Llama3.3-70B-Turbo_003_memory_b000_t00_e02-8d639e26
gaia
Llama3.3-70B-Turbo
2
action
[ "misalignment" ]
["At step 2, the agent's plan was to find and access the first episode of Game Grumps' Sonic the Hed(...TRUNCATED)
[ "action" ]
60
2,591,227
"{\"messages\": [{\"role\": \"user\", \"content\": \"\\nYou are an expert research assistant capable(...TRUNCATED)
"[{\"step\": 2, \"action\": {\"failure_type\": \"misalignment\", \"reasoning\": \"At step 2, the age(...TRUNCATED)
GPT-4o_027_alfworld_task_027
alfworld
GPT-4o
3
memory
[ "over_simplification" ]
["At step 3 the memory module began to oversimplify the agent's exploration history by summarizing t(...TRUNCATED)
[ "memory" ]
60
178,260
"{\"messages\": [{\"role\": \"user\", \"content\": \"\\nYou are an expert agent operating in the ALF(...TRUNCATED)
"[{\"step\": 3, \"memory\": {\"failure_type\": \"over_simplification\", \"reasoning\": \"At step 3 t(...TRUNCATED)
GPT-4o_003_chat_b000_t00_e02-1e119842
webshop
GPT-4o
1
system
[ "environment_error" ]
["The terms ‘man’ and ‘woman’ are conflicting in this context, making the search unlikely to(...TRUNCATED)
[ "system" ]
60
168,839
"{\"messages\": [{\"role\": \"user\", \"content\": \"\\nYou are an expert agent operating in the Web(...TRUNCATED)
"[{\"step\": 1, \"system\": {\"failure_type\": \"environment_error\", \"reasoning\": \"The terms \\u(...TRUNCATED)
Llama3.3-70B-Turbo_006_006
alfworld
Llama3.3-70B-Turbo
7
plan
[ "" ]
[ "" ]
[ "plan" ]
60
212,707
"{\"messages\": [{\"role\": \"user\", \"content\": \"\\nYou are an expert agent operating in the ALF(...TRUNCATED)
[{"step": 7, "plan": {"failure_type": "", "reasoning": ""}}]
GPT-4o_016_alfworld_task_016
alfworld
GPT-4o
30
system
[ "step_limit" ]
[ "In its last step, it had already found and taken the book, but the step limit was reached." ]
[ "system" ]
60
172,312
"{\"messages\": [{\"role\": \"user\", \"content\": \"\\nYou are an expert agent operating in the ALF(...TRUNCATED)
"[{\"step\": 30, \"system\": {\"failure_type\": \"step_limit\", \"reasoning\": \"In its last step, i(...TRUNCATED)
GPT-4o_055_alfworld_task_055
alfworld
GPT-4o
6
reflection
[ "progress_misjudge" ]
["reflection concludes the task is done after placing the bowl on the shelf, assuming it is already (...TRUNCATED)
[ "reflection" ]
60
209,005
"{\"messages\": [{\"role\": \"user\", \"content\": \"\\nYou are an expert agent operating in the ALF(...TRUNCATED)
"[{\"step\": 6, \"reflection\": {\"failure_type\": \"progress_misjudge\", \"reasoning\": \"reflectio(...TRUNCATED)
Llama3.3-70B-Turbo_011_webshop_task_011
webshop
Llama3.3-70B-Turbo
4
plan
[ "plan_inefficient" ]
[ "" ]
[ "plan" ]
60
183,394
"{\"messages\": [{\"role\": \"user\", \"content\": \"\\nYou are an expert agent operating in the Web(...TRUNCATED)
[{"step": 4, "plan": {"failure_type": "plan_inefficient", "reasoning": ""}}]
GPT-4o_052_alfworld_task_052
alfworld
GPT-4o
21
plan
[ "inefficient_plan" ]
["already confirming at steps 3-4 that fridge 1 contains no tomato, the plan again sends the agent b(...TRUNCATED)
[ "plan" ]
60
222,236
"{\"messages\": [{\"role\": \"user\", \"content\": \"\\nYou are an expert agent operating in the ALF(...TRUNCATED)
"[{\"step\": 21, \"plan\": {\"failure_type\": \"inefficient_plan\", \"reasoning\": \"already confirm(...TRUNCATED)
GPT-4o_019_alfworld_task_019
alfworld
GPT-4o
7
action
[ "misalignment" ]
["The plan clearly stated to determine if it was clean, but the action was to take it directly. This(...TRUNCATED)
[ "action" ]
60
244,139
"{\"messages\": [{\"role\": \"user\", \"content\": \"\\nYou are an expert agent operating in the ALF(...TRUNCATED)
"[{\"step\": 7, \"action\": {\"failure_type\": \"misalignment\", \"reasoning\": \"The plan clearly s(...TRUNCATED)
GPT-4o_026_alfworld_task_026
alfworld
GPT-4o
5
reflection
[ "outcome_misinterpretation" ]
["The agent tried to go to fridge 1 but it thought the action was not executed on the observation mo(...TRUNCATED)
[ "reflection" ]
60
185,551
"{\"messages\": [{\"role\": \"user\", \"content\": \"\\nYou are an expert agent operating in the ALF(...TRUNCATED)
"[{\"step\": 5, \"reflection\": {\"failure_type\": \"outcome_misinterpretation\", \"reasoning\": \"T(...TRUNCATED)
End of preview. Expand in Data Studio

davide221/agenterrorbench(複製)

データの内容

エージェントの軌跡 200 件に、決定的な失敗のステップ(critical_failure_step)・モジュール(critical_failure_module)・失敗の種類(failure_types)の列を持つデータ。1 行が 1 件の軌跡で、全体の記録は full_trajectory、ステップごとの注釈は step_annotations にあり、ラベルの出所は未確認である。元のカードは空で、この説明は列名と調査メモから書いた。

データの概要

項目 値
行数 200(1 構成・3 分割の合計)
列数 12
容量 18.26 MB(ファイル 5 本)
ファイルの形式 parquet 3
公開 public

構成と分割

構成(config) 分割(split)ごとの行数 列数
default train 139、validation 31、test 30 12

列(構成 default・分割 train)

列 型 先頭の行の値
trajectory_id string Llama3.3-70B-Turbo_003_memory_b000_t00_e02-8d639…
task_type string gaia
llm_model string Llama3.3-70B-Turbo
critical_failure_step int64 2
critical_failure_module string action
failure_types list<string> [1 件]
failure_reasonings list<string> ["At step 2, the agent's plan was to find and ac…
failure_modules list<string> [1 件]
num_steps int64 60
trajectory_length int64 2591227
full_trajectory string "{\"messages\": [{\"role\": \"user\", \"content\…
step_annotations string "[{\"step\": 2, \"action\": {\"failure_type\": \…

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

davide221/agenterrorbench をリビジョン ac71bfa93c9743319060debb0d2278e02b00640f 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

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: D(記載なし・other・unknown・資料間の食い違い)。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 (L91) API に license 無し 確認済 原文注記: カードが無く、ラベルの出所が未確認。full_trajectory が 1 件で 2,591,227 文字ある行がある(先頭行)。全件が失敗軌跡と読める(列に成功の印が無い)

元のカード

(元のカードは空だった)

Downloads last month
98