mzio/aprm-sft_thinkact-Eaprm_tw_tr_medium_hi-Gaprm_qwen3_ap-S42-R1_ns4_alo0-ap1-train_all-b040

Act-PRM Rollout Dataset

Run Metadata

• —run_name: act-prm-v2-ia=0-rr=0-ec=act_prm_tw_treasure_medium_hi-gc=aprm_qwen3_ap-tc=aprm_for_sft100-rbc=default-alo=0-mc=hf_qwen3_4b_inst_2507-lc=r8_a16_qkvo-nf=0-ao=1-ho=1-gs=8-bs=8-lr=4e_05-ns=4-s=42-r=1_ns4_alo0
• —run_cmd: main_pytorch.py --env_config act_prm/tw_treasure_medium --model_config hf_qwen3_4b_inst_2507 --lora_config r8_a16_qkvo --generator_config aprm_qwen3_ap --trainer_config aprm_for_sft100 --replay_buffer_config default --log_path ./logs --actions_only --hide_observations --batch_size 8 --group_size 8 --learning_rate 4e-5 --num_substeps 1 --seed 42 --replicate 0 --verbose --env_config act_prm/tw_treasure_medium_hi --num_substeps 4 --replicate 1_ns4_alo0 --act_logprob_offset 0
• —batch_idx: 40
• —split: train_all
• —trajectory_key: think_act_policy
• —num_trajectories: 244
• —num_episodes: 244
• —group_size: 8
• —batch_size: 8