{ "model": "EleutherAI/pythia-1b", "n_layers": 16, "L_crit": 15, "n_prompts": 30, "seed": 42, "logp_gap": 1.96565416653951, "mean_delta_pre_lcrit": 42.40066298224842, "cross_wins_layers": 14, "interpretation": "Crystallized residual IS useful when injected early", "per_layer": [ { "layer": 0, "R_std": 0.0, "R_cross": 58.92, "delta": 58.92 }, { "layer": 1, "R_std": 8.25, "R_cross": 48.52, "delta": 40.27 }, { "layer": 2, "R_std": 5.5, "R_cross": 56.91, "delta": 51.41 }, { "layer": 3, "R_std": -1.27, "R_cross": 62.16, "delta": 63.43 }, { "layer": 4, "R_std": -2.09, "R_cross": 65.77, "delta": 67.86 }, { "layer": 5, "R_std": 2.64, "R_cross": 64.96, "delta": 62.32 }, { "layer": 6, "R_std": 8.85, "R_cross": 65.87, "delta": 57.02 }, { "layer": 7, "R_std": 9.39, "R_cross": 65.24, "delta": 55.86 }, { "layer": 8, "R_std": 23.36, "R_cross": 63.94, "delta": 40.58 }, { "layer": 9, "R_std": 25.04, "R_cross": 66.4, "delta": 41.36 }, { "layer": 10, "R_std": 31.25, "R_cross": 64.8, "delta": 33.54 }, { "layer": 11, "R_std": 32.09, "R_cross": 66.88, "delta": 34.79 }, { "layer": 12, "R_std": 51.02, "R_cross": 68.28, "delta": 17.26 }, { "layer": 13, "R_std": 69.72, "R_cross": 72.96, "delta": 3.23 }, { "layer": 14, "R_std": 73.34, "R_cross": 81.51, "delta": 8.16 }, { "layer": 15, "R_std": 92.53, "R_cross": 92.53, "delta": 0.0 } ] }