Views
No views yet
onnx/model_quantized.onnx — int8 quantized weights (preferred for browser)config.json / ort_config.json / inference_config.json1{
2 "threshold": 0.55,
3 "thresholds": {
4 "1": 0.55,
5 "2": 0.55,
6 "3": 0.55
7 },
8 "base_model": "allenai/scibert_scivocab_uncased",
9 "max_len": 96,
10 "num_epochs": 3,
11 "labels": [
12 "none",
13 "numerical_ambiguity",
14 "publication_issue",
15 "novelty_issue"
16 ],
17 "test_macro_precision": 0.9955631447299272,
18 "test_macro_recall": 0.9966948148948124,
19 "test_macro_f1": 0.996124676084053,
20 "test_issue_precision": 0.995229642777901,
21 "test_issue_recall": 0.9983307367015357,
22 "test_issue_f1": 0.9967777777777778,
23 "test_issue_auc": 0.9994281464031068,
24 "test_per_class_auc": {
25 "numerical_ambiguity": 0.9999369309842113,
26 "publication_issue": 0.9999273817182374,
27 "novelty_issue": 0.9999737853019018
28 },
29 "n_train": 90000,
30 "n_val": 15000,
31 "n_test": 15000,
32 "train_minutes": 77.012782116731,
33 "hard_eval": {
34 "slice": "quality_hard_eval.jsonl (12,386 rows: fresh unarXive blobs 400k-650k, unseen paraphrase templates, hard negatives)",
35 "any_issue_precision": 0.989,
36 "any_issue_recall": 0.677,
37 "any_issue_f1": 0.804,
38 "fp_rate_natural_clean": 0.009,
39 "fp_rate_hard_negatives": 0.021,
40 "recall_natural_weak_label_positives": 0.988,
41 "recall_unseen_template_positives": 0.443,
42 "note": "Weak-label test metrics (~0.996 F1) overstate quality; this harder slice is the honest reference. Precision-first: false flags are rare, but recall drops on phrasings unlike the training templates."
43 }
44}