Datasets:
Agentic Trajectory Recall v3 (ENERZAi 내부, 1차 업로드 2026-09-22)
실제 에이전트 궤적(neulab/agent-data-collection 표준화본: nebius SWE-agent, swe-play, swe-gym openhands, openhands, AgentTuning alfworld/db/kg/os/webshop)을
AMA-Bench compaction_v3_nostate 하네스 형식(Task + Step Index + Most Recent + Recalled steps + Questions + Answer[N]:)의 단일 user 메시지로 렌더링하고,
궤적에서 프로그램으로 정답을 뽑은 질문 15종(전사·탐색·집계·관계·증거부재)을 붙인 학습/검증 데이터. 삼진(W1.58) Qwen3-1.7B 의 장기 기록 회상 학습용.
AMA-Bench 테스트 원문은 포함하지 않는다 — 형식만 차용. WebArena 계열 소스(go-browse-wa, nnetnav-wa)는 AMA WEB 과 같은 사이트라 제외. 에피소드별 AMA 테스트셋 8-gram 겹침 15% 초과 행 제거(최대 8.97%), 질문/정답 누출 0.
구성(config) — 각 config 의 split 은 출처(source) 이다
| config | 내용 | 행 |
|---|---|---|
short_nothink |
A 팔 학습 데이터: 짧은 입력(정답 근거 10~20단계 창), think 없음. 12,892행 | 12,892 |
short_teacher |
배치5 학습 데이터: 짧은 입력 + Qwen3.8-27B 교사 think(react-일치 채택). 10,213행 | 10,213 |
long_nothink |
긴 입력(~6K 토큰) 결함 수정판(v3), think 없음. 12,002행 — 학습엔 아직 미사용 | 12,002 |
long_teacher |
B 팔 학습 데이터: 긴 입력 + 27B 교사 think. 7,950행 | 7,950 |
holdout_short |
짧은 검증 문제지(학습 미사용 궤적, md5(id)%10==0). 460행/2,965문항 | 460 |
holdout_long |
긴 검증 문제지(결함 수정판). 454행/2,846문항 | 454 |
long_v1_defective |
배치3 학습 데이터(결함 포함 v1: 답 없는 질문·숫자 먼저 형식 등). 재현용. 12,282행 | 12,282 |
holdout_v1_defective |
v1 검증 문제지(결함 포함). 456행 | 456 |
학습 결과(같은 출발점 배치3 에서 6천만 토큰씩 이어 학습; 짧은 검증 / 긴 검증 문항 정확도 / AMA, 모두 think 끔):
short_nothink → 94.9 / 84.0 / 16.15 · short_teacher → 93.2 / 81.1 / 17.11 · long_teacher → 89.9 / 87.4 / 17.23 · long_v1_defective(배치3) → 81.2 / 84.6 / 17.43.
공식 Qwen3-1.7B AMA think 끔 27.92. 상세는 학습 레포 data_gen/longctx/DECISIONS.md.
열(column)
| 열 | 형 | 뜻 |
|---|---|---|
id |
str | 궤적 id(concat 모드는 <첫id>+<n>x<k>) |
source |
str | real/<출처> |
messages |
str(JSON) | [{role:user, content:<AMA 형식 프롬프트>}, {role:assistant, content:"Answer[1]: … Answer[N]: …", reasoning_content?: <교사 think>}] |
questions / answers / qtypes |
list[str] | 질문·프로그램 정답·질문 유형(15종) |
evidence_steps |
list[list[int]] | 각 질문의 근거 단계 번호 |
keys / kinds |
list[list[str]] / list[str] | 규칙 채점용 정답 키·채점 종류(text/step/count/bool/tokens/refuse) |
answer_format |
str | indexed(Answer[N]) 또는 natural |
meta |
str(JSON) | {concat, n_steps, prompt_tokens, traj_chars} — 문자열로 직렬화(원본은 dict) |
think_source / think_tokens / teacher_preds |
str / int / str(JSON) | 교사 think 출처(teacher27b_react_match)·길이·교사가 낸 답(문자열 직렬화). teacher config 만 |
교사 think: Qwen3.8-27B-FP8 이 정답을 보지 않고 <reasoning> 을 쓰고 답을 냈을 때 모든 답이 프로그램 정답과 일치한 시도만 채택(최대 3회, T 0.3/0.7/0.7, think ≤512 토큰). 채택률 짧은 79.2% · 긴 66.2%.
학습 yaml 예
- {name: HBKenerzai/agentic-recall-real-v3, subset: long_teacher, split: nebius_SWE_agent_trajectories, group: real,
require_thinking: true, max_think_tokens: 640, think_inline: true, transform: longctx/recall_v1}
think 없는 config 는 require_thinking: false — 트레이너가 <think>\n\n</think>\n\n 를 채워 think 끔 서빙과 바이트 동일하게 렌더링한다.
알려진 한계
- 질문이 단계 번호를 지목한 전사·세기 위주(정답 p50 9단어)라, AMA 가 묻는 개체 단위 상태 변화·원인·가정(정답 p50 38단어, 38% 가 여러 단계 종합)은 재지 않는다. 검증 문제지 이득이 AMA 로 옮겨가지 않았다(6팔 반복 확인).
long_v1_defective는 답이 문맥에 없는 질문(단어 단계 유형 47.8%)·숫자 먼저 형식 등 결함이 있다. 재현 외 용도로 쓰지 말 것.
- Downloads last month
- 13