Add step100 LoRA adapters for WebShop FFN Channel-SAM (7B and 1.5B)
#3
by Davidk1 - opened
Intermediate step-100 checkpoints for the two completed WebShop FFN Channel-SAM runs (rho=0.005, max_abs_delta=0.05, last layer, layerwise RMS, activation/BF16 gain, clean rollout, seed 0).
LoRA adapter + tokenizer/config only; FSDP model/optimizer shards are not included.
Internal 128-goal validation at step100 (not the official 500-goal evaluation):
Qwen2.5-7B-Instruct: success_rate 0.594, task_score 0.738 (step200: 0.695 / 0.801)
Qwen2.5-1.5B-Instruct: success_rate 0.539, task_score 0.753 (step200: 0.664 / 0.822)
Each run's existing run_config.json on main documents its settings; the 7B entry additionally records host-specific deviations and the step-16 actor micro-batch amendment.