| parameters: | |
| project_name: jaxgmg2_3phase_optim_state_alpha1 | |
| action: rl | |
| rl_action: train | |
| # Learning — alpha changed to 1.0 from base model's 0.6 | |
| lr: 5e-5 | |
| alpha: 1.0 | |
| discount_rate: 0.98 | |
| cheese_loc: any | |
| env_layout: open | |
| mask_type: first_episode | |
| use_prev_action: false | |
| # Training scale | |
| num_total_env_steps: 7_372_800_000 # 12_000 grad steps * 64 rollout steps * 9600 levels | |
| num_levels: 9600 | |
| grad_acc_per_chunk: 4 | |
| num_rollout_steps: 64 | |
| # Resume from jaxgmg2_3phase_optim_state checkpoint 3810, with optimizer state | |
| resume: jaxgmg2_3phase_optim_state/al_0.6_g_0.98_id_17_seed_980617 | |
| resume_id: 3810 | |
| resume_optim: true | |
| # Checkpointing | |
| ckpt_dir: jaxgmg2_3phase_optim_state_alpha1 | |
| checkpoint: al_0.6_g_0.98_id_17_seed_980617_resume_alpha1 | |
| eval_schedule: "0:1,250:2,500:5,2000:10" | |
| log_optimizer_state: true | |
| deterministic: true | |
| # Logging | |
| use_wandb: true | |
| use_hf: true | |
| hf_user: urdshals | |
| wandb_project: jaxgmg2_patt | |
| sweep: | |
| - seed: 42 | |