v3.20.0

ruvnet/ruflov3.20.0Jul 4, 2026by ruvnet

AI Summary

Enhances the native training flywheel with validation splitting, resume capabilities, and checkpoint auto-load, requiring ruvllm 2.6.0.

Key Highlights

Validation split with early stopping and bestValLoss tracking
Resume capability to continue training from a checkpoint
Checkpoints auto-load on first adaptation use
Bumped @ruvector/ruvllm to 2.6.0 for new metadata and geometry validation

New Features

Validation split (--val-split)
Resume training (--resume)
Checkpoint auto-load
Native pipeline resume support

Full Release Notes

Native training pipeline gains real epochs-across-runs and closes the train→checkpoint→better-routing loop.

- `neural train --val-split <frac>` — validation + early stopping (surfaces bestValLoss/earlyStopped)
- `neural train --resume <checkpoint>` — continues from the restored epoch via ruvllm 2.6.0 resumeFrom() (degrades to 2.5.7 weight-restore)
- **Checkpoint auto-load** — routing's lazy LoRA adapter loads the newest checkpoint on first adaptation use (off the startup hot path, kill-switch, non-fatal)
- `neural status` shows latest checkpoint + age
- **@ruvector/ruvllm 2.6.0** (RuVector#638): checkpoint v2 metadata + geometry validation, true resumeFrom(), best-checkpoint retention; floors bumped >=2.6.0

Tests 12/12 · startup 0.08s (no regression) · PR #2557