{ "benchmark": "calibrated Aygiz-derived bilingual evaluation sample", "evaluation_rows": 30000, "positive_rows": 10000, "negative_rows": 20000, "device": "cuda", "results": [ { "model": "nano_scorer", "roc_auc": 0.9497601, "average_precision": 0.878473879740799, "seconds": 4.01, "pairs_per_second": 7472.2 }, { "model": "devlake_bert", "roc_auc": 0.98955255, "average_precision": 0.9706141532855449, "seconds": 266.76, "pairs_per_second": 112.5 }, { "model": "labse_shallow_cosine", "roc_auc": 0.9899108474999999, "average_precision": 0.9773271326270535, "seconds": 57.87, "pairs_per_second": 518.4 }, { "model": "medium_scorer", "roc_auc": 0.9633081225, "average_precision": 0.9155399783751561, "seconds": 11.74, "pairs_per_second": 2555.9 }, { "model": "labse_cosine", "roc_auc": 0.9944407325, "average_precision": 0.9894755664187669, "seconds": 669.0, "pairs_per_second": 44.8 }, { "model": "laser_cosine", "roc_auc": 0.979622155, "average_precision": 0.9596781034776389, "seconds": 669.0, "pairs_per_second": 44.8 }, { "model": "labse_laser_mean", "roc_auc": 0.993943075, "average_precision": 0.9871495585646881, "seconds": 669.0, "pairs_per_second": 44.8 }, { "model": "mini_scorer_fresh", "roc_auc": 0.9590461200000001, "average_precision": 0.9040391252158363, "seconds": 6.42, "pairs_per_second": 4673.5 } ], "reference_note": "The evaluation includes a dedicated LaBSE + LASER baseline run; full-corpus scoring artifacts are not part of this public release." }