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
..
2026-04-11 11:54:09 +02:00
2026-04-02 16:29:55 +02:00
2026-03-14 11:35:15 +01:00
2026-03-13 10:18:35 +01:00
2026-03-13 10:18:35 +01:00
2026-04-15 07:01:19 +02:00
2026-03-13 10:18:35 +01:00
2026-03-13 10:18:35 +01:00
2026-03-13 10:18:35 +01:00
2026-04-13 17:09:34 +02:00
2026-03-13 10:18:35 +01:00
2026-04-13 19:28:00 +02:00
2026-03-29 22:35:37 +02:00
2026-03-29 22:35:37 +02:00
2026-04-08 21:31:46 +02:00
2026-03-29 19:18:47 +02:00
2026-03-19 00:39:03 +01:00
2026-03-13 10:18:35 +01:00
2026-03-13 10:18:35 +01:00
2026-04-13 17:09:34 +02:00
2026-03-16 21:01:28 +01:00
2026-03-13 10:18:35 +01:00
2026-03-13 10:18:35 +01:00
2026-03-13 10:18:35 +01:00
2026-04-10 18:54:14 +02:00
2026-03-13 10:18:35 +01:00
2026-03-13 10:18:35 +01:00
2026-03-13 10:18:35 +01:00
2026-03-18 00:53:47 +01:00
2026-03-14 11:35:15 +01:00
2026-03-15 11:59:31 +01:00