PRM-agent-rl-artifacts

Training rollouts and evaluation outputs for the GRPO agent runs in this project. Each rollout file is one optimizer step; each line is one sampled trajectory with its decoded prompt, response and reward.

Contents

• rollouts/search_r1_qwen3_8b_4gpu — Search-R1 / Qwen3-8B outcome-GRPO training rollouts
• rollouts/search_r1_qwen3_8b_perturnnorm — Search-R1 / Qwen3-8B Process-GRPO (per-turn-norm) training rollouts
• rollouts/alfworld_qwen3_8b_gigpo — ALFWorld / Qwen3-8B GiGPO training rollouts, one JSONL per optimizer step
• rollouts/webshop_qwen3_8b_gigpo — WebShop / Qwen3-8B GiGPO training rollouts, one JSONL per optimizer step
• rollouts/webshop_olmo3_7b_gigpo — WebShop / OLMo-3-7B GiGPO training rollouts, one JSONL per optimizer step
• rollouts/webshop_olmo3_7b_outcome_2048 — WebShop / OLMo-3-7B outcome-GRPO (2048 cap) training rollouts, one JSONL per optimizer step
• rollouts/webshop_qwen3_8b_process_center_mean_std — WebShop / Qwen3-8B Process-GRPO (center mean_std) training rollouts, abandoned at step 124
• rollouts/webshop_qwen3_8b_process_std_sum — WebShop / Qwen3-8B Process-GRPO (std sum) training rollouts, abandoned at step 75
• eval/standalone_eval — Trainer-free checkpoint evaluations: per-step metrics, per-episode trajectories, aggregates
• eval/search_r1_metrics — Search-R1 per-checkpoint evaluation metrics and per-sample outputs