Exa model eval metrics
Browse files
production_model/eval_metrics_exa_v1.json
ADDED
|
@@ -0,0 +1,136 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model": "reranker_exa_v1",
|
| 3 |
+
"data": "93K Exa findSimilar pairs + hard negatives",
|
| 4 |
+
"train_rows": 39459,
|
| 5 |
+
"eval_rows": 3881,
|
| 6 |
+
"train_queries": 2257,
|
| 7 |
+
"eval_queries": 235,
|
| 8 |
+
"best_iteration": 188,
|
| 9 |
+
"eval_metrics": {
|
| 10 |
+
"ndcg@10": 0.999861211859234,
|
| 11 |
+
"hard_neg_auc": 0.9730603385975726,
|
| 12 |
+
"mrr": 0.9659574468085106
|
| 13 |
+
},
|
| 14 |
+
"importance": [
|
| 15 |
+
{
|
| 16 |
+
"f": "position_inverse",
|
| 17 |
+
"i": 91858.95870601013
|
| 18 |
+
},
|
| 19 |
+
{
|
| 20 |
+
"f": "candidate_position",
|
| 21 |
+
"i": 14132.449180345982
|
| 22 |
+
},
|
| 23 |
+
{
|
| 24 |
+
"f": "co_citation_count",
|
| 25 |
+
"i": 3560.862382028252
|
| 26 |
+
},
|
| 27 |
+
{
|
| 28 |
+
"f": "query_citation_count",
|
| 29 |
+
"i": 1034.9953849189915
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"f": "candidate_influential_citations",
|
| 33 |
+
"i": 937.888320078142
|
| 34 |
+
},
|
| 35 |
+
{
|
| 36 |
+
"f": "age_ratio",
|
| 37 |
+
"i": 918.4462667144835
|
| 38 |
+
},
|
| 39 |
+
{
|
| 40 |
+
"f": "citation_count_ratio",
|
| 41 |
+
"i": 836.8483927757479
|
| 42 |
+
},
|
| 43 |
+
{
|
| 44 |
+
"f": "query_age_days",
|
| 45 |
+
"i": 783.7682531564496
|
| 46 |
+
},
|
| 47 |
+
{
|
| 48 |
+
"f": "category_x_recency",
|
| 49 |
+
"i": 772.5697443634272
|
| 50 |
+
},
|
| 51 |
+
{
|
| 52 |
+
"f": "candidate_num_cited_by",
|
| 53 |
+
"i": 769.316386341583
|
| 54 |
+
},
|
| 55 |
+
{
|
| 56 |
+
"f": "candidate_citation_count",
|
| 57 |
+
"i": 722.9314735834487
|
| 58 |
+
},
|
| 59 |
+
{
|
| 60 |
+
"f": "shared_author_count",
|
| 61 |
+
"i": 651.7318598488346
|
| 62 |
+
},
|
| 63 |
+
{
|
| 64 |
+
"f": "citations_x_recency",
|
| 65 |
+
"i": 589.5717827547342
|
| 66 |
+
},
|
| 67 |
+
{
|
| 68 |
+
"f": "candidate_citations_per_year",
|
| 69 |
+
"i": 467.35710863163695
|
| 70 |
+
},
|
| 71 |
+
{
|
| 72 |
+
"f": "candidate_age_days",
|
| 73 |
+
"i": 375.18497193418443
|
| 74 |
+
},
|
| 75 |
+
{
|
| 76 |
+
"f": "candidate_recency_score",
|
| 77 |
+
"i": 242.9321630364284
|
| 78 |
+
},
|
| 79 |
+
{
|
| 80 |
+
"f": "query_num_references",
|
| 81 |
+
"i": 179.99511690903455
|
| 82 |
+
},
|
| 83 |
+
{
|
| 84 |
+
"f": "year_diff",
|
| 85 |
+
"i": 103.86553149763495
|
| 86 |
+
},
|
| 87 |
+
{
|
| 88 |
+
"f": "query_log_citations",
|
| 89 |
+
"i": 80.68094129022211
|
| 90 |
+
},
|
| 91 |
+
{
|
| 92 |
+
"f": "same_primary_category",
|
| 93 |
+
"i": 72.33195366058499
|
| 94 |
+
}
|
| 95 |
+
],
|
| 96 |
+
"production_ready": true,
|
| 97 |
+
"schema": [
|
| 98 |
+
"qdrant_cosine_score",
|
| 99 |
+
"candidate_position",
|
| 100 |
+
"candidate_citation_count",
|
| 101 |
+
"candidate_log_citations",
|
| 102 |
+
"candidate_influential_citations",
|
| 103 |
+
"candidate_age_days",
|
| 104 |
+
"candidate_recency_score",
|
| 105 |
+
"query_citation_count",
|
| 106 |
+
"query_age_days",
|
| 107 |
+
"year_diff",
|
| 108 |
+
"same_primary_category",
|
| 109 |
+
"co_citation_count",
|
| 110 |
+
"shared_author_count",
|
| 111 |
+
"candidate_is_newer",
|
| 112 |
+
"query_log_citations",
|
| 113 |
+
"citation_count_ratio",
|
| 114 |
+
"age_ratio",
|
| 115 |
+
"candidate_citations_per_year",
|
| 116 |
+
"query_num_references",
|
| 117 |
+
"candidate_num_cited_by",
|
| 118 |
+
"ewma_longterm_similarity",
|
| 119 |
+
"ewma_shortterm_similarity",
|
| 120 |
+
"ewma_negative_similarity",
|
| 121 |
+
"cluster_importance",
|
| 122 |
+
"cluster_distance_to_medoid",
|
| 123 |
+
"is_suppressed_category",
|
| 124 |
+
"onboarding_category_match",
|
| 125 |
+
"user_total_saves",
|
| 126 |
+
"user_total_dismissals",
|
| 127 |
+
"user_days_since_last_save",
|
| 128 |
+
"user_session_save_count",
|
| 129 |
+
"cosine_x_recency",
|
| 130 |
+
"cosine_x_citations",
|
| 131 |
+
"category_x_recency",
|
| 132 |
+
"cosine_x_cocitation",
|
| 133 |
+
"position_inverse",
|
| 134 |
+
"citations_x_recency"
|
| 135 |
+
]
|
| 136 |
+
}
|