Revision Assistant — Manuscript Quality

Quantized ONNX + tokenizer for browser / Transformers.js inference in Revision Assistant.

Files

• onnx/model_quantized.onnx — int8 quantized weights (preferred for browser)
• tokenizer + config.json / ort_config.json / inference_config.json

Metrics / inference config

{
  "threshold": 0.55,
  "thresholds": {
    "1": 0.55,
    "2": 0.55,
    "3": 0.55
  },
  "base_model": "allenai/scibert_scivocab_uncased",
  "max_len": 96,
  "num_epochs": 3,
  "labels": [
    "none",
    "numerical_ambiguity",
    "publication_issue",
    "novelty_issue"
  ],
  "test_macro_precision": 0.9955631447299272,
  "test_macro_recall": 0.9966948148948124,
  "test_macro_f1": 0.996124676084053,
  "test_issue_precision": 0.995229642777901,
  "test_issue_recall": 0.9983307367015357,
  "test_issue_f1": 0.9967777777777778,
  "test_issue_auc": 0.9994281464031068,
  "test_per_class_auc": {
    "numerical_ambiguity": 0.9999369309842113,
    "publication_issue": 0.9999273817182374,
    "novelty_issue": 0.9999737853019018
  },
  "n_train": 90000,
  "n_val": 15000,
  "n_test": 15000,
  "train_minutes": 77.012782116731,
  "hard_eval": {
    "slice": "quality_hard_eval.jsonl (12,386 rows: fresh unarXive blobs 400k-650k, unseen paraphrase templates, hard negatives)",
    "any_issue_precision": 0.989,
    "any_issue_recall": 0.677,
    "any_issue_f1": 0.804,
    "fp_rate_natural_clean": 0.009,
    "fp_rate_hard_negatives": 0.021,
    "recall_natural_weak_label_positives": 0.988,
    "recall_unseen_template_positives": 0.443,
    "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."
  }
}