File size: 2,259 Bytes
cf7c2d8 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 | ---
library_name: bibr
base_model: sentence-transformers/all-MiniLM-L6-v2
language: en
tags:
- text-classification
- scholarly-documents
---
# bibr section classifier
Two-head classifier that bibr uses to type section headings it cannot resolve
with its alias lookup. It predicts one of 16 canonical section types and
whether the heading is top level. Encoder: all-MiniLM-L6-v2 (22M parameters),
mean pooling, input template v3 (heading, position bucket, neighbouring
headings, start of the body).
This revision retrains the same architecture and template on a larger,
re-labelled corpus. It is a drop-in replacement for the September 2026
release: same files, same label order, same runtime cost.
## Training data
- `section_v5b_leakfiltered`: section headings from PMC Open Access JATS
papers (English), paper-grouped train/val/test split. Three papers removed
by a later leak re-check are excluded.
- Train subsample of at most 3,000 rows per class: 45,979 rows from 6,946 papers.
Epoch chosen on the v5b validation split only.
- No bibr-validation gold, blind evaluation, or held-out benchmark paper is in
the training or validation data (matched by DOI, PMCID and title).
## Evaluation
Canonical-type macro-F1 / accuracy (model-routed rows: macro-F1). "Model-routed" is the subset bibr
actually sends to this model (headings the alias lookup misses), scored on
bibr's section gold.
| set | previous release | this revision |
|---|---|---|
| v5b test (English JATS, 9,507 headings) | 0.589 / 0.398 | 0.966 / 0.946 |
| gold, English headings (9,501) | 0.656 / 0.572 | 0.798 / 0.784 |
| gold, non-English headings (1,524) | 0.213 / 0.295 | 0.265 / 0.344 |
| gold, model-routed English (3,472) | 0.419 | 0.676 |
| gold, model-routed non-English (1,240) | 0.168 | 0.217 |
Part of the gain on gold comes from shared labelling conventions: the
training data and the gold both label subsections by their parent type and
structured-abstract parts as abstract.
Non-English headings are zero-shot: the training data is English only.
## Files
`model.safetensors`, `tokenizer.json`, `tokenizer_config.json`,
`label_classes.json`, `inference_config.json` (template version 3), and an
ONNX Runtime bundle in `onnx/` used by bibr's core install.
|