| { |
| "_comment": "Test-time loop-scaling on Sudoku-Extreme. Train depth L_cycles=n_backwards_L=6; eval at scaled depths. exact-puzzle (seq) acc, EMA weights, 12 batches x 768 = 9216 puzzles. eval_loopscale.py.", |
| "depths": [6, 9, 12, 18, 24], |
| "attn_host": { |
| "base": {"s0": [0.5934, 0.4646, 0.6246, 0.5355, 0.5160], |
| "s1": [0.5945, 0.6041, 0.6007, 0.6086, 0.6117]}, |
| "loop_attn": {"s0": [0.6009, 0.5558, 0.1198, 0.0075, 0.0034], |
| "s1": [0.6318, 0.6534, 0.6681, 0.6826, 0.6943], |
| "s2": [0.5632, 0.5726, 0.5816, 0.5846, 0.5882]} |
| }, |
| "mlp_host": { |
| "base": {"s0": [0.8457, 0.8492, 0.8544, 0.8564, 0.8624], |
| "s1": [0.8403, 0.8468, 0.8478, 0.8535, 0.8614], |
| "s2": [0.8502, 0.8535, 0.8621, 0.8733, 0.8762]}, |
| "loop_attn": {"s0": [0.8508, 0.8503, 0.8681, 0.8749, 0.8796], |
| "s1": [0.8528, 0.8700, 0.8877, 0.8978, 0.9092], |
| "s2": [0.8457, 0.8545, 0.8684, 0.8894, 0.9034]} |
| }, |
| "mlp_host_means": { |
| "base": {"6": 0.8454, "9": 0.8498, "12": 0.8548, "18": 0.8611, "24": 0.8667}, |
| "loop_attn": {"6": 0.8498, "9": 0.8583, "12": 0.8747, "18": 0.8874, "24": 0.8974}, |
| "loop_attn_minus_base": {"6": 0.004, "9": 0.009, "12": 0.020, "18": 0.026, "24": 0.031} |
| } |
| } |
|
|
