Add qwen2_5_1_5_stable_gaussian_noise_v9_step_200_seed0 step200 LoRA adapter and run config
Add completed WebShop Qwen2.5-1.5B stable Gaussian v9, step200, seed0.
Preflight: latest checkpoint=200; nonempty step200 LoRA adapter; sigma=0.2, alpha=0.0; warmup=20; global_scale_mode=prepass_grad_squared; global_scale_clip_multiplier=2.0; statistics_timing=post_update; target_global_step=200. Training log reports exit status 0. The original run_config.json retains its launch-time status="active"; it is uploaded unchanged for provenance.
Only add files under results/webshop/robust_training/qwen2_5_1_5_stable_gaussian_noise_v9_step_200_seed0/: run_config.json, step200 actor/lora_adapter, and TensorBoard events. No optimizer/model shards, raw rollouts, tar archives, deletions, or modifications to existing main files. LoRA rank64 adapter is 295,488,904 bytes (not a full checkpoint shard).
OOM recovery: both runs changed PPO_MICRO from8 to4, with all other training hyperparameters unchanged. v9 restarted from pretrained step0; v11 resumed checkpoint step5 (first5 steps used micro8, subsequent updates used micro4). TensorBoard retains original and resumed attempts.
Requested v13 (sigma0.4, alpha1.0): its local run directory was not found in the inspected checkouts, so no v13 artifacts are included; current training step cannot be confirmed. v9 and v11 are submitted in separate PRs. Existing PR #3 remains open and is not modified by this upload.