Spaces:
Sleeping
Sleeping
File size: 1,338 Bytes
ad7798f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 | {
"purpose": "Product-identity-disjoint evaluation of real multi-view product pairs",
"seed": 20260917,
"primary_metric": "binary F1 on fixed 1:2 positive:negative sampling",
"threshold_grid": {"minimum": 0, "maximum": 100, "step": 0.5},
"profile_selection": "Calibration only; maximize F1, then precision, then recall, then prefer baseline profile. Freeze before test scoring.",
"production_profiles": {
"current": [0.30, 0.30, 0.20, 0.20],
"local_emphasis": [0.30, 0.50, 0.10, 0.10],
"hash_emphasis": [0.60, 0.30, 0.05, 0.05],
"structure_emphasis": [0.20, 0.20, 0.50, 0.10],
"equal_weights": [0.25, 0.25, 0.25, 0.25]
},
"component_order": ["perceptual", "features", "ssim", "color"],
"ablation_profiles": {"hash_only": [1,0,0,0], "features_only": [0,1,0,0], "ssim_only": [0,0,1,0], "color_only": [0,0,0,1]},
"baseline_threshold": 75,
"test_policy": "No test-driven tuning. Report baseline and frozen selected profile even if the candidate does not improve.",
"latency_policy": "Server compute timing is distinct from client end-to-end latency. Zero-weight rescoring is accuracy ablation only, not an algorithm latency benchmark.",
"abstention_policy": "Insufficient-texture decisions are reported separately and treated as nonmatches for end-to-end binary metrics. Also report coverage."
}
|