maze-transformer / experiments /eval_sft_ref.json

feat: unified experiment launcher — every experiment = one JSON config (experiments/*.json); launch.py dispatches rl_nav/rl_rnn/rl_ctx/evaluate/random; 11 configs + 11 tests

ec6c2c6 3 days ago

Raw

History Blame Contribute Delete

337 Bytes

	{
	"name": "eval_sft_ref",
	"task": "evaluate",
	"note": "Q1 对照：SFT 参考权重（监督 2000 步）在 forced-obs 协议下的表现",
	"checkpoint": "checkpoints/l2_d64_maze_sft_ref.pt",
	"model_type": "gpt",
	"n_trials": 24,
	"seed": 7,
	"budget_scale": 2.0,
	"sizes": ["5x5", "6x6", "7x7"],
	"include_random": true
	}