Input: the embedding caches of stage 03. One GPU and about 80 GB of host memory each; the training split of a cache is held in memory.
python -m clin_jepa.training.train_predictor --config configs/train/predictor_vjepa2ac.yaml
python -m clin_jepa.training.train_predictor --config configs/train/predictor_sft_baseline.yaml
python -m clin_jepa.training.train_predictor --config configs/train/predictor_separate.yaml| model | cache | output |
|---|---|---|
| V-JEPA 2-AC style | embeddings/vjepa2ac |
predictor_vjepa2ac/best.pt |
| SFT baseline w/o JEPA | embeddings/sft_baseline |
predictor_sft_baseline/best.pt |
| Clin-JEPA w/ separately trained predictor | embeddings/clin_jepa |
predictor_separate/seed_{42,1337,2026}/best.pt |
Twenty epochs (15,400 steps) each, under an hour per training.