bibr-front-role-v1

A small gradient-boosted classifier that assigns a role to every text region of a scientific paper's page: title, byline, affiliation, abstract, keywords, doi_line, masthead, heading, ref_header, body, other.

Used by bibr as evidence alongside its lexical front-matter heuristics — never as a decision on its own. It reads page-relative geometry, relative font size and script-independent text shape, so it recognises a byline the English name-shape test cannot read.

Trust boundary

front_role.joblib is a pickle. bibr loads it through a gadget-restricted loader, but you should still only point ML_FRONT_ROLE_MODEL_ID at a checkpoint you trust.

Training data

Publisher JATS projected onto cached OCR regions, so the labels are verbatim ground truth rather than an LLM's opinion. 328,781 training rows.

Metrics

Held-out test split, 19,252 regions: accuracy 0.955, macro F1 0.893.

role	precision	recall	F1	support
title	1.000	0.964	0.981	110
doi_line	0.994	0.981	0.987	876
byline	0.922	0.801	0.858	267
affiliation	0.891	0.946	0.918	241
abstract	0.963	0.706	0.815	296
keywords	0.934	0.486	0.640	146
ref_header	0.833	0.798	0.815	94
heading	0.982	0.989	0.985	2121
masthead	0.927	0.871	0.898	1285
body	0.955	0.973	0.964	5321
other	0.951	0.969	0.960	8495

Per paper (106 test papers): the top-scoring title region is a true title region on 100%, and predicted byline regions cover the true byline on 96%.

Effect in bibr

Replayed over 192 gold-annotated papers, comparing front-matter resolution with and without the model on identical page contents:

	title	byline	abstained
heuristics alone	0.849	0.260	0.104
+ this model	0.901	0.698	0.047

10 titles fixed and 0 broken; 84 bylines fixed and 0 broken.

License

Apache-2.0.