diff options
Diffstat (limited to 'artifacts/manifold_gate/confirm_vg_k128_seed1.json')
| -rw-r--r-- | artifacts/manifold_gate/confirm_vg_k128_seed1.json | 476 |
1 files changed, 476 insertions, 0 deletions
diff --git a/artifacts/manifold_gate/confirm_vg_k128_seed1.json b/artifacts/manifold_gate/confirm_vg_k128_seed1.json new file mode 100644 index 0000000..8af213b --- /dev/null +++ b/artifacts/manifold_gate/confirm_vg_k128_seed1.json @@ -0,0 +1,476 @@ +{ + "protocol": "Content-projected states, directions fitted without pairs on held-out scenes; the complete assignment gate is scored with hidden pairs.", + "dataset": "vg", + "dims": 128, + "samples": 512, + "gate_a": { + "true": { + "mse": 0.8461898686561814, + "mse_channels": [ + 0.8461898686561814 + ], + "mse_channel_mean": 0.8461898686561814, + "conditional_kl": 83.11779233731765, + "m30_total": 18.315938204775893 + }, + "similarity_shift": { + "mse": 1.909032153101283, + "mse_channels": [ + 1.909032153101283 + ], + "mse_channel_mean": 1.909032153101283, + "conditional_kl": 270.97437517635547, + "m30_total": 58.01293934147366 + }, + "random": { + "mse": { + "mean": 1.9999308586120605, + "std": 0.005415276624262333, + "min": 1.980623722076416, + "true_z": 213.05300903320312, + "true_rank_among_random": 0, + "count": 1000 + }, + "mse_channel_mean": { + "mean": 1.9999308586120605, + "std": 0.005415276624262333, + "min": 1.980623722076416, + "true_z": 213.05300903320312, + "true_rank_among_random": 0, + "count": 1000 + }, + "conditional_kl": { + "mean": 281.0531311035156, + "std": 2.571009874343872, + "min": 272.8770446777344, + "true_z": 76.98738861083984, + "true_rank_among_random": 0, + "count": 1000 + }, + "m30_total": { + "mean": 60.21048355102539, + "std": 0.5166908502578735, + "min": 58.56594467163086, + "true_z": 81.0824203491211, + "true_rank_among_random": 0, + "count": 1000 + } + } + }, + "gate_b": { + "pairs": 130816, + "improving_pairs": 369, + "improving_fraction": 0.0028207558708414873, + "delta_mean": 0.008980280748833325, + "delta_min": -0.0026841194923196954, + "identity_is_local_min_mse": false, + "worst_improving_swaps": [ + { + "pair": [ + 421, + 501 + ], + "delta_mse": -0.0026841194923196954, + "text_cosine": -0.15778705247305966, + "visual_cosine": -0.9528240803641709 + }, + { + "pair": [ + 97, + 358 + ], + "delta_mse": -0.002663336340810483, + "text_cosine": 1.824243065556641, + "visual_cosine": -1.2295060553249777 + }, + { + "pair": [ + 169, + 253 + ], + "delta_mse": -0.0019376819915811598, + "text_cosine": 0.7411414924722992, + "visual_cosine": -0.12430072148946003 + }, + { + "pair": [ + 358, + 501 + ], + "delta_mse": -0.001759521451084722, + "text_cosine": -0.4570102430219021, + "visual_cosine": 0.09365572381683304 + }, + { + "pair": [ + 495, + 504 + ], + "delta_mse": -0.0017329221887820603, + "text_cosine": 1.8233645577922106, + "visual_cosine": 1.8167843883128862 + }, + { + "pair": [ + 290, + 362 + ], + "delta_mse": -0.0017143033905624486, + "text_cosine": 4.833723360149344, + "visual_cosine": 2.8130833861304616 + }, + { + "pair": [ + 340, + 504 + ], + "delta_mse": -0.0016158181222582677, + "text_cosine": 0.44165668487704335, + "visual_cosine": 0.48974504750913267 + }, + { + "pair": [ + 236, + 290 + ], + "delta_mse": -0.001608743744623828, + "text_cosine": 4.402974780275268, + "visual_cosine": 3.0131067462252985 + }, + { + "pair": [ + 169, + 501 + ], + "delta_mse": -0.001558727521990226, + "text_cosine": 0.37290694783608724, + "visual_cosine": -0.7707538415646742 + }, + { + "pair": [ + 483, + 495 + ], + "delta_mse": -0.0014727686838593296, + "text_cosine": 2.6100475839823387, + "visual_cosine": 2.027944087295041 + }, + { + "pair": [ + 290, + 467 + ], + "delta_mse": -0.001469326474868847, + "text_cosine": 2.8110644714236526, + "visual_cosine": 2.764876905132282 + }, + { + "pair": [ + 253, + 403 + ], + "delta_mse": -0.0014280543862510368, + "text_cosine": 2.6343809883613627, + "visual_cosine": 1.6839976736528914 + }, + { + "pair": [ + 386, + 453 + ], + "delta_mse": -0.0014036385435340988, + "text_cosine": 2.9330325452513204, + "visual_cosine": 3.1382799042932943 + }, + { + "pair": [ + 354, + 495 + ], + "delta_mse": -0.0013316803569251613, + "text_cosine": 2.857800917910961, + "visual_cosine": 2.0846964322364956 + }, + { + "pair": [ + 358, + 495 + ], + "delta_mse": -0.0012986005965767281, + "text_cosine": 2.343889544997752, + "visual_cosine": 1.8531494155174117 + }, + { + "pair": [ + 97, + 504 + ], + "delta_mse": -0.0012463199691556785, + "text_cosine": 0.7387633361913704, + "visual_cosine": -0.7025267926765577 + }, + { + "pair": [ + 231, + 246 + ], + "delta_mse": -0.0012053636206535458, + "text_cosine": 3.514154434836688, + "visual_cosine": 1.1216217867064673 + }, + { + "pair": [ + 105, + 501 + ], + "delta_mse": -0.0011685985778248956, + "text_cosine": 2.6505501917922656, + "visual_cosine": 1.0674686580497288 + }, + { + "pair": [ + 358, + 426 + ], + "delta_mse": -0.0011073270400286717, + "text_cosine": 1.680278852913856, + "visual_cosine": 0.5305612100173487 + }, + { + "pair": [ + 437, + 501 + ], + "delta_mse": -0.001106646981753433, + "text_cosine": 0.08057126708720055, + "visual_cosine": 1.0510485545739516 + } + ] + }, + "descent_from_true": { + "start_accuracy": 1.0, + "final_accuracy": 0.521484375, + "final_objective": 0.7794426591572278, + "final_energy": { + "mse": 0.7794426591572278, + "mse_channels": [ + 0.7794426591572278 + ], + "mse_channel_mean": 0.7794426591572278, + "conditional_kl": 78.03057734468294, + "m30_total": 17.165000787251042 + }, + "accepted_moves": 201, + "moved_fraction": 0.478515625, + "mse_local_min_certificate": true, + "trajectory": [ + { + "step": 0, + "objective": 0.8435057491638618, + "accuracy": 0.99609375 + }, + { + "step": 50, + "objective": 0.8020735811325613, + "accuracy": 0.833984375 + }, + { + "step": 100, + "objective": 0.7891407447994455, + "accuracy": 0.708984375 + }, + { + "step": 150, + "objective": 0.7821254946263017, + "accuracy": 0.595703125 + }, + { + "step": 200, + "objective": 0.7794426591572278, + "accuracy": 0.521484375 + } + ] + }, + "descent_from_random": [ + { + "start_accuracy": 0.00390625, + "final_accuracy": 0.0078125, + "final_objective": 0.9158073469489326, + "final_energy": { + "mse": 0.9158073469489326, + "mse_channels": [ + 0.9158073469489326 + ], + "mse_channel_mean": 0.9158073469489326, + "conditional_kl": 116.65519837241126, + "m30_total": 25.16265436838012 + }, + "accepted_moves": 779, + "moved_fraction": 0.9921875, + "mse_local_min_certificate": true + }, + { + "start_accuracy": 0.0, + "final_accuracy": 0.001953125, + "final_objective": 1.025276503048015, + "final_energy": { + "mse": 1.025276503048015, + "mse_channels": [ + 1.025276503048015 + ], + "mse_channel_mean": 1.025276503048015, + "conditional_kl": 125.33430178222437, + "m30_total": 27.11741336254091 + }, + "accepted_moves": 785, + "moved_fraction": 0.986328125, + "mse_local_min_certificate": true + }, + { + "start_accuracy": 0.001953125, + "final_accuracy": 0.0234375, + "final_objective": 0.91658106446105, + "final_energy": { + "mse": 0.91658106446105, + "mse_channels": [ + 0.91658106446105 + ], + "mse_channel_mean": 0.91658106446105, + "conditional_kl": 129.0357364536353, + "m30_total": 27.640309419649164 + }, + "accepted_moves": 771, + "moved_fraction": 0.986328125, + "mse_local_min_certificate": true + }, + { + "start_accuracy": 0.0, + "final_accuracy": 0.0, + "final_objective": 1.021352128329811, + "final_energy": { + "mse": 1.021352128329811, + "mse_channels": [ + 1.021352128329811 + ], + "mse_channel_mean": 1.021352128329811, + "conditional_kl": 131.58247432525, + "m30_total": 28.359199121709622 + }, + "accepted_moves": 785, + "moved_fraction": 0.98046875, + "mse_local_min_certificate": true + }, + { + "start_accuracy": 0.001953125, + "final_accuracy": 0.005859375, + "final_objective": 1.0090331149061613, + "final_energy": { + "mse": 1.0090331149061613, + "mse_channels": [ + 1.0090331149061613 + ], + "mse_channel_mean": 1.0090331149061613, + "conditional_kl": 132.22582945848274, + "m30_total": 28.46323212150887 + }, + "accepted_moves": 789, + "moved_fraction": 0.9765625, + "mse_local_min_certificate": true + }, + { + "start_accuracy": 0.0, + "final_accuracy": 0.0390625, + "final_objective": 0.8892596534628971, + "final_energy": { + "mse": 0.8892596534628971, + "mse_channels": [ + 0.8892596534628971 + ], + "mse_channel_mean": 0.8892596534628971, + "conditional_kl": 116.84754948938762, + "m30_total": 25.14802920480332 + }, + "accepted_moves": 793, + "moved_fraction": 0.990234375, + "mse_local_min_certificate": true + }, + { + "start_accuracy": 0.001953125, + "final_accuracy": 0.021484375, + "final_objective": 0.8914677956933827, + "final_energy": { + "mse": 0.8914677956933827, + "mse_channels": [ + 0.8914677956933827 + ], + "mse_channel_mean": 0.8914677956933827, + "conditional_kl": 113.04674933573207, + "m30_total": 24.39228545853318 + }, + "accepted_moves": 834, + "moved_fraction": 0.990234375, + "mse_local_min_certificate": true + }, + { + "start_accuracy": 0.001953125, + "final_accuracy": 0.0, + "final_objective": 0.9680916194838789, + "final_energy": { + "mse": 0.9680916194838789, + "mse_channels": [ + 0.9680916194838789 + ], + "mse_channel_mean": 0.9680916194838789, + "conditional_kl": 125.86645001989251, + "m30_total": 27.10947324294626 + }, + "accepted_moves": 821, + "moved_fraction": 0.990234375, + "mse_local_min_certificate": true + }, + { + "start_accuracy": 0.0, + "final_accuracy": 0.017578125, + "final_objective": 0.9195944096607847, + "final_energy": { + "mse": 0.9195944096607847, + "mse_channels": [ + 0.9195944096607847 + ], + "mse_channel_mean": 0.9195944096607847, + "conditional_kl": 121.97179857463767, + "m30_total": 26.233548534249106 + }, + "accepted_moves": 813, + "moved_fraction": 0.984375, + "mse_local_min_certificate": true + }, + { + "start_accuracy": 0.00390625, + "final_accuracy": 0.0, + "final_objective": 0.9696137786871593, + "final_energy": { + "mse": 0.9696137786871593, + "mse_channels": [ + 0.9696137786871593 + ], + "mse_channel_mean": 0.9696137786871593, + "conditional_kl": 127.62678515447998, + "m30_total": 27.464584588270316 + }, + "accepted_moves": 778, + "moved_fraction": 0.9921875, + "mse_local_min_certificate": true + } + ], + "verdict": { + "true_z_mse": 213.05300903320312, + "improving_fraction": 0.0028207558708414873, + "identity_strict_local_min": false, + "descent_keeps": 0.521484375, + "true_mse": 0.8461898686561814, + "best_random_descent": 0.8892596534628971, + "best_random_accuracy": 0.0390625, + "counterfeit_found": false + } +}
\ No newline at end of file |