Upload evaluation/gold_results.json with huggingface_hub
Browse files- evaluation/gold_results.json +72 -0
evaluation/gold_results.json
ADDED
|
@@ -0,0 +1,72 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"entity_f1": 0.2917,
|
| 3 |
+
"entity_precision": 0.2414,
|
| 4 |
+
"entity_recall": 0.3684,
|
| 5 |
+
"ci_95_low": 0.1519,
|
| 6 |
+
"ci_95_high": 0.4422,
|
| 7 |
+
"relaxed_f1": 0.6557,
|
| 8 |
+
"relaxed_precision": 0.6087,
|
| 9 |
+
"relaxed_recall": 0.7105,
|
| 10 |
+
"relaxed_ci_95_low": 0.5067,
|
| 11 |
+
"relaxed_ci_95_high": 0.7838,
|
| 12 |
+
"per_aspect_relaxed": {
|
| 13 |
+
"QUANTITY": {
|
| 14 |
+
"relaxed_p": 0.4,
|
| 15 |
+
"relaxed_r": 0.4,
|
| 16 |
+
"relaxed_f1": 0.4,
|
| 17 |
+
"n_gold": 5,
|
| 18 |
+
"n_pred": 5
|
| 19 |
+
},
|
| 20 |
+
"SUBSTITUTION": {
|
| 21 |
+
"relaxed_p": 0.4783,
|
| 22 |
+
"relaxed_r": 1.0,
|
| 23 |
+
"relaxed_f1": 0.6471,
|
| 24 |
+
"n_gold": 11,
|
| 25 |
+
"n_pred": 23
|
| 26 |
+
},
|
| 27 |
+
"TECHNIQUE": {
|
| 28 |
+
"relaxed_p": 0.8182,
|
| 29 |
+
"relaxed_r": 0.5714,
|
| 30 |
+
"relaxed_f1": 0.6729,
|
| 31 |
+
"n_gold": 14,
|
| 32 |
+
"n_pred": 11
|
| 33 |
+
},
|
| 34 |
+
"ADDITION": {
|
| 35 |
+
"relaxed_p": 0.8571,
|
| 36 |
+
"relaxed_r": 0.75,
|
| 37 |
+
"relaxed_f1": 0.8,
|
| 38 |
+
"n_gold": 8,
|
| 39 |
+
"n_pred": 7
|
| 40 |
+
}
|
| 41 |
+
},
|
| 42 |
+
"token_f1": 0.4175,
|
| 43 |
+
"token_precision": 0.4511,
|
| 44 |
+
"token_recall": 0.4418,
|
| 45 |
+
"span_error_counts": {
|
| 46 |
+
"false_positive": 17,
|
| 47 |
+
"under_extension": 13,
|
| 48 |
+
"off_by_one": 5,
|
| 49 |
+
"false_negative": 10,
|
| 50 |
+
"exact_match": 14,
|
| 51 |
+
"aspect_confusion": 1,
|
| 52 |
+
"over_extension": 0,
|
| 53 |
+
"prefix_error": 0
|
| 54 |
+
},
|
| 55 |
+
"boundary_errors": 13,
|
| 56 |
+
"pct_off_by_one": 0.3846,
|
| 57 |
+
"pct_prefix_error": 0.0,
|
| 58 |
+
"span_count_stats": {
|
| 59 |
+
"total_gold_spans": 38,
|
| 60 |
+
"total_pred_spans": 46,
|
| 61 |
+
"avg_gold_per_positive": 1.583,
|
| 62 |
+
"avg_pred_per_positive": 1.333,
|
| 63 |
+
"n_positive_examples": 24
|
| 64 |
+
},
|
| 65 |
+
"prefix_examples": [],
|
| 66 |
+
"n_examples": 496,
|
| 67 |
+
"model_path": "models/checkpoints/dictabert_crf/P1_crf_weighted",
|
| 68 |
+
"test_file": "data/gold_validation/gold_tokenized_dictabert_v2.jsonl",
|
| 69 |
+
"classification_report": " precision recall f1-score support\n\n ADDITION 0.38 0.38 0.38 8\n QUANTITY 0.20 0.20 0.20 5\nSUBSTITUTION 0.26 0.73 0.38 11\n TECHNIQUE 0.14 0.14 0.14 14\n\n micro avg 0.24 0.37 0.29 38\n macro avg 0.24 0.36 0.27 38\nweighted avg 0.23 0.37 0.27 38\n",
|
| 70 |
+
"gold_scheme": "BIO",
|
| 71 |
+
"model_scheme": "BIO"
|
| 72 |
+
}
|