Files
foxhunt/crates
jgrusewski 1fdc712df6 feat: trajectory backtracking — checkpoint rewind + informed perturbation
Detects Q-gap plateau (5 consecutive frozen epochs), rewinds to
best-improvement checkpoint, applies perturbation strategy cycle:
1. Reset Adam momentum
2. Shrink-perturb branch heads
3. Temperature boost (2× for 5 epochs)
4. Learning rate boost (2× for 10 epochs)

3 rewinds × 4 routes = 12 max attempts. On exhaustion: save best model,
exit with PLATEAU_EXHAUSTED. Saves ~3 hours H100 vs frozen training.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-15 07:01:19 +02:00
..