菜单
Diagnoses onset event F1, train/eval matching mismatches, confidence collapse, and threshold effects. Use when F1 is zero or low, conv1d collapse is suspected, or the user asks about Hungarian vs ordered assignment, matching tolerance, or debug_onset_overfit output.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Runs StepCOVNet GPU training and MERT extraction on Windows via WSL with TensorFlow CUDA libraries. Use for train_onset, overfit scripts, extract_mert_features with cuda, or when TensorFlow sees zero GPUs in WSL.
One-prompt agent autoresearch loop for StepCOVNet: goal + time budget → orient, hypothesize, run, log, repeat until success or deadline. Use when the user says "run autoresearch", gives a research goal with a budget, or names a profile (e.g. ar-tide-overfit). Not for unattended Python planners (--hours / lattice).
Deprecated alias — use autoresearch skill with profile ar-tide-overfit.
Starts a research session and logs EXP and NOTE entries with timestamps. Use when beginning research work, logging experiments, or writing discussion notes. Promote findings to PAPER_OUTLINE only when drafting the paper.
Audit and sync agent brain — read all rules from disk, reconcile skills/catalogs/routing. Triggers on refresh agent brain, refresh all rules, audit rules, or after steering promotion. Read-only audit script; agent writes doc updates.
Triggers on process mistakes and steering corrections; run steering-correction-promotion for artifact choice, context optimization, and JRN receipt. Journal alone is incomplete.
| name | onset-event-eval-matching |
| description | Diagnoses onset event F1, train/eval matching mismatches, confidence collapse, and threshold effects. Use when F1 is zero or low, conv1d collapse is suspected, or the user asks about Hungarian vs ordered assignment, matching tolerance, or debug_onset_overfit output. |
| disable-model-invocation | true |
Training and eval must use the same assignment philosophy. Eval always uses Hungarian matching with tolerance; training uses assign_onset_pairs_l1 (Hungarian L1) in src/stepcovnet/onset_events/losses.py.
Ordered slot→GT pairing on tide (634 GT, 1024 uniform slots) can yield zero in-tolerance pairs → loss drives all confidences toward 0 → 0% F1 despite reasonable predicted times (see NOTE-20260606-13, EXP-07/08).
Before changing epochs, architecture, or loss weights, run diagnostics via the WSL command template with:
scripts/debug_onset_overfit.py --checkpoint=models_wsl/overfit_tide/<frontend>/onset_event_model.keras
Check in output / diagnostics.py:
| Symptom | Likely cause | Next step |
|---|---|---|
| Max conf ~0.0002, 0% F1 | Train/eval assignment mismatch or no learnable pairs | Verify Hungarian train loss; read NOTE-13 |
| Many pairs in tolerance, low F1 | Threshold / post-hoc filtering | tide-ablations threshold phase |
| ~28–30% F1, high FP | Query-slot formulation cap, not metric bug | Formulation change; see EXP-10 |
| Module | Role |
|---|---|
losses.py | assign_onset_pairs_l1 — training assignment |
metrics.py | match_onsets_numpy — eval Hungarian matching |
diagnostics.py | Confidence sweeps, pair stats |
scripts/debug_onset_overfit.py | Checkpoint inspection CLI |
Log insight as NOTE-… if design-relevant; EXP-… if a fix was validated by retrain.
