Aug 27 18:45 webshop_single_expert_qwen2_5_3_step_300_seed0: add best_val official eval (0.590/0.732) 18:45 webshop_single_expert_qwen2_5_3_step_300_seed0: add best_val official eval (0.590/0.732) 18:21 webshop_single_expert_qwen2_5_3_step_300_seed0: LoRA adapters (steps 100/200/300, best_val) + manifest/README; first 3B run, below 1.5B reference Aug 26 03:24 Remove pre-fix and legacy-layout WebShop results Remove sampler-pre-fix runs, obsolete full checkpoints, and flat-layout duplicates after canonical sampler-fixed artifacts were uploaded and verified. 03:23 Publish canonical sampler-fixed WebShop training checkpoints Mirror the local results/webshop/training contract with pruning-ready LoRA and sanitized router artifacts. No pruning outputs or FSDP shards. Aug 24 21:20 Add sampler-fixed WebShop step-300 single- and 2-expert checkpoints Publish compact PEFT inference checkpoints for the sampler-fixed 300-step single-expert and per-expert-sample-matched 2-expert WebShop experiments. The 2-expert router state is sanitized to remove rollout/diversity buffers. Aug 19 19:58 Add 2expert clusterF nothinking final (step1000) 19:58 Add 2expert clusterF nothinking best (step500, val SR 0.742) 19:58 Add 1expert nothinking final (step1000) 19:57 Add 1expert nothinking best (step980, val SR 0.805) 19:57 Add 2expert mlp random: best==last==step1000 (val SR 0.758) Jun 21 16:58 add single-expert global_step_400 Jun 19 19:41 add single-expert best_val (step620 succ0.828) 19:37 add single-expert global_step_1000 19:26 initial commit 
