Distinct-kink complexity regularizer and merger / results/metrics.json
Failed on benchmark
1{
2 "device": "cuda",
3 "seed": 559,
4 "train_seconds": 2.990062952041626,
5 "math_algebra_max_error": 1.6689300537109375e-06,
6 "constructed_width": 48,
7 "merged_width": 24,
8 "width_reduction_factor": 2.0,
9 "probe_max_abs_function_error": 1.430511474609375e-06,
10 "probe_rmse_function_error": 2.1097835656291863e-07,
11 "val_mse_redundant": 0.019573688507080078,
12 "val_mse_merged": 0.01957368478178978,
13 "val_mse_compact": 0.01957368291914463,
14 "jacobian_mean_norm_redundant": 1.3525495529174805,
15 "jacobian_mean_norm_merged": 1.3525495529174805,
16 "params_redundant": 147,
17 "params_merged": 75,
18 "note": "The redundant model is an exact coefficient-split reparameterization of a trained compact model; this isolates the claimed merger phenomenon rather than claiming spontaneous duplicate discovery."
19}