{ "name": "eval_selfplay_ref", "task": "evaluate", "note": "Q2 对照:废弃的 selfplay 权重(自产 obs 训练)在 forced-obs 协议下的表现", "checkpoint": "checkpoints/l2_d64_maze_selfplay_final.pt", "model_type": "gpt", "n_trials": 24, "seed": 7, "budget_scale": 2.0, "sizes": ["5x5", "6x6", "7x7"], "include_random": true }