PengxinWang

/

RobustLLMAgent

Like 2

WebShop 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).