Commit History
-
Add WebShop 7B single-layer FFN Channel-SAM step200 (rho=0.005, cap=0.05)
Verified signatureWebShop GiGPO + single-layer FFN Channel-SAM, Qwen2.5-7B-Instruct, seed 0, fresh start to step 200. Intervention: last-layer FFN intermediate channels (18944 channels), rho=0.005, max_abs_delta=0.05, layerwise RMS normalization, activation (BF16) gain arithmetic, clean rollout. LoRA adapter + tokenizer/config only; FSDP model/optimizer shards are not included. Internal 128-goal validation at step200: success_rate 0.695, task_score 0.801 (not the official 500-goal evaluation). Run-host deviations from the published 7B baselines are recorded in run_config.json: tensor_model_parallel_size=2, rollout load_format=safetensors, layered_summon=True, actor param_offload=True, and actor micro-batch 8->4 from step 16 (see provenance_amendments).9bba64e
-
Publish ALFWorld Qwen2.5-7B step200 clean-vs-gaussian evaluation results
Verified signature15aff1e -
Publish qwen2_5_7_stable_gaussian_noise_step_200_seed0 step200 LoRA adapter for collaborator ALFWorld 7B handoff
Verified signature230907a -
Publish qwen2_5_7_stable_gaussian_noise_step_200_seed0 step200 LoRA adapter for collaborator ALFWorld 7B handoff
Verified signatureb30a182 -
Publish qwen2_5_7_stable_gaussian_noise_step_200_seed0 step200 LoRA adapter for collaborator ALFWorld 7B handoff
Verified signaturee87bc1e -
Publish qwen2_5_7_stable_gaussian_noise_step_200_seed0 step200 LoRA adapter for collaborator ALFWorld 7B handoff
Verified signatureda23783
-
Remove redundant motivation README
Verified signature707f7f1 -
Remove redundant reward surface README
Verified signature56454c7 -
Organize reward surface and motivation figures with reproducible CSV inputs
Verified signaturebbaca61
-
Publish complete WebShop step200 three-seed evaluation results
Verified signature230ae6e -
Remove superseded SAM-named evaluation figure
Verified signaturef322c84 -
Publish complete WebShop step200 three-seed evaluation results
Verified signature5ad165d
-
Use fresh cap 0.05 SAM as official step200 baseline
Verified signatureb756470 -
Add WebShop single-layer FFN Channel-SAM step200 (rho=0.005, cap=0.05) (#1)
- Add WebShop single-layer FFN Channel-SAM step200 (rho=0.005, cap=0.05) (2be24292d326dab2fee8258dd1149b4dc5f8226a) Co-authored-by: Daivd Li <Davidk1@users.noreply.huggingface.co>aa7d451 -
Remove redundant evaluation package readme and manifest
Verified signaturec8c5c49 -
Update curated evaluation figures and reproducible data
Verified signatureadbb800 -
Use Vanilla, Plain Gaussian and Stable Gaussian algorithm labels
Verified signature057000e -
Publish WebShop 1.5B Stable Gaussian step200 main result
Verified signature9d5f63d -
Simplify model card around step200 baseline and model download table
Verified signature3cdd8a6
-
Document unified 200-step training default for 1.5B and 7B
Verified signatureSet the documented default to step200 while preserving historical step300 artifact paths and accurately describing published weights. Documentation only; no checkpoint changes.b562238
-
Publish four completed Qwen2.5-7B step-200 baseline adapters
Verified signatureALFWorld/WebShop clean and Gaussian baselines, seed 0; terminal LoRA adapters, original run configs, SHA256 manifests and loading instructions. Existing 1.5B artifacts retained.2f9d4a3
-
Document four finalized step-300 evaluation baselines
Verified signature7903637 -
Publish evaluation-only step-300 LoRA baselines; remove training archive
Verified signature903ca99 -
Archive baseline: alfworld/training/qwen2_5_1_5_step_300_seed0
Verified signature5dfb70c
-
initial commit
Verified signature6e62342