File size: 428 Bytes
4435a09 | 1 2 3 4 5 | # ablate-5 stage2 — 1 EPOCH (s2ep1)
Monet sft_stage2 on ablate-5 (5 cats: bbox_crop/depth/edge/helper_interleaved/scene_graph, ~40k=39998), from base Qwen2.5-VL-7B.
**stage2 trained 1 epoch** (retrained fresh, NOT a 2-epoch midpoint ckpt — LR fully decays over 313 steps). latent=8, ce4.0/al2.0/emph2.0, eff batch 128.
Used to precompute ablate-5 stage3 teacher latents (-> ulvr_subset/ablate-5/stage3_s2ep1). Weights only.
|