tide-ablations

name	tide-ablations
description	Runs tide overfit ablations — confidence threshold sweep, loss-weight variants, and architecture changes on MERT. Use when investigating F1 plateau, loss weight tuning, threshold post-hoc, or arch_medium vs arch_large on tide.
disable-model-invocation	true

Tide ablations

Script

scripts/run_overfit_tide_ablations.py

Use the WSL command template with:

scripts/run_overfit_tide_ablations.py --epochs=50 --phase=all

Phases:

Phase	What it does
threshold	Sweep confidence on existing MERT checkpoint (no retrain)
train	Retrain MERT with loss-weight and architecture variants
all	Threshold sweep then training ablations (default)

Output

Artifact	Path
Summary	models_wsl/overfit_tide_ablations/ablation_summary.json
Per-run models	models_wsl/overfit_tide_ablations/<variant>/
Callbacks	callbacks/overfit_tide_ablations/<variant>/

EXP-10 conclusions (do not re-run blindly)

• Threshold sweep: no meaningful F1 lift — recall capped before post-hoc threshold helps
• Loss weights (cls-heavy, time-heavy, both-high): ~same ~28–30% F1
• arch_medium (embed 128): similar to baseline
• arch_large (embed 256): OOM on 8 GB GPU — skip or reduce batch

Plateau is not fixed by epochs, loss weights, threshold, or medium arch bump. Likely query-slot formulation (high FP, ~233/634 recall).

After the run

1. Read ablation_summary.json (if script crashed, check terminal + partial models)
2. Log EXP-… in EXPERIMENT_LOG.md; tag pipeline stages: train for loss, model for arch, post for threshold
3. Next: multi-song val or formulation alternatives per EXPERIMENT_LOG.md § Recommended next step

Known pitfall

sweep_confidence_thresholds requires numeric min_onset_distance_ms (use 0.0, not None). See JRN-20260606-05 in self-journal.md.