review-log
Read an entire runs/<id>/log.md and assess research narrative coherence, identify stuck hypotheses, detect overfitting signs, and write a review markdown. Use after a full run completes or when a loop has been stuck for many rounds.
소스 정보
- 저장소
- qualit527/qec-ai-decoder
- 최근 소스 활동
- 2026년 4월 22일 05:58
- 감지된 SKILL.md 언어
- 영어
- 스타
- 3
- 포크
- 0
설치 방법
기본적으로 소스를 먼저 확인하는 Prompt가 선택됩니다. 직접 명령으로 전환하거나 로컬 사본을 다운로드할 수도 있습니다.
소스 파일 검토
설치 여부를 결정하기 전에 SKILL.md와 SkillsMP에 표시된 보조 파일을 읽어 보세요.
SKILL.md 표시 중
SKILL.md
소스 지침 · 읽기 전용 미리보기- name
- review-log
- description
- Read an entire runs/<id>/log.md and assess research narrative coherence, identify stuck hypotheses, detect overfitting signs, and write a review markdown. Use after a full run completes or when a loop has been stuck for many rounds.
# /review-log
## When to use
- A run of 10+ rounds has completed.
- User asks for a "research review" or "is the agent stuck?"
## Inputs
- `run_dir`: path to `runs/<id>/`
## Behavior
1. Run `python -m cli.autoqec review-log <run_dir>` to get structured stats (round count, Pareto size, top hypotheses, killed_by_safety count).
2. Read `log.md` fully.
3. LLM-reason:
- Narrative coherence: does each round build on previous findings?
- Stuck patterns: ≥3 rounds with near-identical hypotheses?
- Overfitting signs: Δ_LER monotonically improving on train seeds but not holdout?
- Safety-kill clustering: are VRAM or wall-clock kills bunched around one hypothesis family?
4. Write `runs/<id>/review.md` with: summary (5 sentences), top 3 concerns, recommended next actions.
## Output
- Path to `review.md`.
GitHub에서 보기