heron-physics-pdf-layout

A fine-tune of docling-project/docling-layout-heron (RT-DETRv2) specialized for layout detection on physics PDFs — articles, ICRC conference proceedings, theses, and book chapters. Same 17-class DocLayNet/Heron label schema and classification head as the base model; only the weights are tuned.

Training data

377 pages from 40 physics-domain PDFs, hand-labeled via a two-model annotation protocol (one model annotates, a second independently judg adjudicated) and split by document — 249 train / 61 val / 67 test pages, no document crossing a split.

Training procedure

transformers.Trainer, AdamW, 20 epochs, batch size 2, tw (backbone lr 1e-5, rest of the model lr 1e-4), horizontal flip augmentation, no replay data mixed in from the base model's original traini

Evaluation (test split, same pages scored against the u

model	mAP@[.5:.95]	AP@0.5	footnote AP	footnote F
base (docling-layout-heron)	0.705	0.835	0.741*
this model	0.841	0.894	0.659	0.873

*the base model's footnote AP looks competitive but is misleading — 21.4% precision means 4 of 5 footnote calls are wrong; it only posts a dec high across the whole precision-recall curve. footnote F1 is the fairer comparison.

Intended use and limitations

Scoped to physics-domain PDFs specifically — not evaluated on, and not claimed to generalize to, general document layout. Part of the Physics-LLM Project.

License

Apache-2.0, inherited from the base model. Weights only; no training code or data redistributed here.