File size: 512 Bytes
4435a09 | 1 2 3 4 5 6 7 8 | # ablate-3 stage3 models (3-category ~40k ablation)
Monet no-text stage3, trained from base Qwen2.5-VL-7B on ablate-3 teacher latents (ulvr_subset/ablate-3/stage3).
Config: stage2=2epoch, stage3=2epoch, latent=8, ce4.0/al2.0.
- **checkpoint-313 = epoch1 = BEST** (epoch2 overfits on the narrow 3-cat data)
- checkpoint-626 = epoch2 (V*/MMVP/BLINK/MMStar all drop vs ep1)
Eval (10 benches) in eval_results §6.8/§6.10; preds in eval_results_before_all/ablate-3-ckpt313|626.
Weights only (optimizer/rng omitted).
|