Spaces:
Sleeping
Sleeping
| { | |
| "dataset": "wiki", | |
| "questions": 600, | |
| "candidate_count_per_question": 10, | |
| "selection_rule": "same neural_evidence_graph_core.graph_arrays and graph_select as four new datasets: 0.4 * per-query minmax relevance + 0.5 * max selected-to-candidate directed title-body link + 0.2 * query-title match; no tuning, no query facets, no specialized presets", | |
| "embedding_source": "previously frozen/precomputed actual all-MiniLM-L6-v2 q and passage vectors; cosine recomputed", | |
| "crossencoder_source": "previously frozen/precomputed actual pretrained ms-marco-MiniLM-L6-v2 256-wordpiece query-passage logits, same ids", | |
| "complete_evidence_coverage_pct": { | |
| "dense_cosine": { | |
| "2": 24.83, | |
| "3": 34.17, | |
| "5": 52.33, | |
| "8": 81.5 | |
| }, | |
| "crossencoder": { | |
| "2": 33.83, | |
| "3": 44.5, | |
| "5": 58.67, | |
| "8": 79.33 | |
| }, | |
| "dense_graph": { | |
| "2": 61.5, | |
| "3": 68.33, | |
| "5": 93.0, | |
| "8": 98.33 | |
| }, | |
| "crossencoder_graph": { | |
| "2": 62.83, | |
| "3": 70.17, | |
| "5": 93.17, | |
| "8": 98.17 | |
| } | |
| }, | |
| "crossencoder_graph_vs_crossencoder_at3": { | |
| "diff_pp": 25.67, | |
| "ci95_pp": [ | |
| 22.17, | |
| 29.33 | |
| ], | |
| "wins": 159, | |
| "losses": 5, | |
| "ties": 436 | |
| }, | |
| "dense_graph_vs_dense_at3": { | |
| "diff_pp": 34.17, | |
| "ci95_pp": [ | |
| 30.0, | |
| 38.0 | |
| ], | |
| "wins": 209, | |
| "losses": 4, | |
| "ties": 387 | |
| }, | |
| "graph_link_present_pct": 81.67, | |
| "gold_support_count_distribution": { | |
| "2": 459, | |
| "4": 141 | |
| }, | |
| "timing": { | |
| "graph_feature_extraction_and_four_rankings_ms_per_question": 0.322, | |
| "graph_feature_extraction_and_four_rankings_total_s": 0.1933, | |
| "pretrained_crossencoder_scoring_ms_per_question_from_original_inference_run": 254.92, | |
| "cached_embedding_and_crossencoder_scores": true | |
| }, | |
| "note": "This is an actual rerun of ranking/selection and evidence coverage from cached neural scores; scoring step was previously executed, not repeated. Scores are on the same selected 600 held-out question IDs." | |
| } |