---
library_name: pytorch
tags:
  - bibr
  - reference-parsing
  - token-classification
  - crf
  - modernbert
---

# bibr Parser v4.5 Gold

ModernBERT + feature-gated CRF reference-field parser for bibr's
`REF_PARSE_STRATEGY=ner` path.

## Training

- Encoder: `answerdotai/ModernBERT-base`
- Initial checkpoint: `parser_giant_v4_fixed/best.pt`
- Fine-tune corpus: `data/parser_v4_5_gold`
- Train rows: 16,459
- Validation rows: 4,031
- Tags: 39 BIO parser tags
- Max sequence length: 256
- Tokenizer convention: `add_special_tokens=False`

## Validation

Best training validation micro entity F1: `0.9653892730`.

Same-split compatibility check against bibr's runtime all-zero token features:

| Checkpoint | Feature Mode | Validation Entity F1 |
| --- | ---: | ---: |
| `parser_giant_v4_fixed/best.pt` | zero features | `0.0543892926` |
| this checkpoint | true features | `0.9651912333` |
| this checkpoint | zero features | `0.9652572991` |
| this checkpoint | no features (`None`) | `0.8276018305` |

bibr's runtime parser should continue passing all-zero `token_features`, not
`None`, because the feature projection bias is part of the trained forward path.

## Files

- `best.pt` - PyTorch state dict for `FeatureGatedEncoderCRF`
- `metrics.json` - training validation summary
- `report.txt` - per-field validation classification report
- `tag_config.json` - tag metadata from `bibr_training.tags`