Upload evaluation/silver_results.json with huggingface_hub
Browse files
evaluation/silver_results.json
ADDED
|
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"entity_f1": 0.3007,
|
| 3 |
+
"entity_precision": 0.2722,
|
| 4 |
+
"entity_recall": 0.3358,
|
| 5 |
+
"ci_95_low": 0.2357,
|
| 6 |
+
"ci_95_high": 0.3709,
|
| 7 |
+
"relaxed_f1": 0.5524,
|
| 8 |
+
"relaxed_precision": 0.5913,
|
| 9 |
+
"relaxed_recall": 0.5182,
|
| 10 |
+
"token_f1": 0.462,
|
| 11 |
+
"token_precision": 0.4902,
|
| 12 |
+
"token_recall": 0.4545,
|
| 13 |
+
"n_examples": 1518,
|
| 14 |
+
"model_path": "models/checkpoints/dictabert_crf/P1_crf_weighted",
|
| 15 |
+
"test_file": "data/processed_v2/test.jsonl",
|
| 16 |
+
"classification_report": " precision recall f1-score support\n\n ADDITION 0.12 0.14 0.13 14\n QUANTITY 0.27 0.29 0.28 21\nSUBSTITUTION 0.27 0.38 0.32 48\n TECHNIQUE 0.31 0.37 0.34 54\n\n micro avg 0.27 0.34 0.30 137\n macro avg 0.24 0.29 0.27 137\nweighted avg 0.27 0.34 0.30 137\n",
|
| 17 |
+
"gold_scheme": "BIO",
|
| 18 |
+
"model_scheme": "BIO"
|
| 19 |
+
}
|