citeparse-32m

citeparse-32m labels the parts of a reference string: authors, title, container, date, volume, issue, pages, DOI and 8 more fields. It is the default model of the citeparse package, which turns the labels into CSL-JSON or BibTeX.

pip install citeparse
citeparse refs.txt --bibtex

Results

700 real references from Crossref, each scored against the cited work's own Crossref record. None of these DOIs were in training. GROBID was given the same DOI rule as citeparse.

Parser All fields References fully right Title Container
GROBID 0.8.1 (CRF) 0.911 71.1% 0.882 0.805
citeparse-32m 0.939 80.3% 0.921 0.858
citeparse-150m 0.952 83.4% 0.934 0.878

GROBID wins on its own held-out citations: 0.943 against 0.882 for this model, on GROBID's own scorer. Those come from the same sources as GROBID's training data.

Speed on CPU (8 threads): about 118 references a second.

Training

From jhu-clsp/ettin-encoder-32m, two epochs of BIO token classification over 16 labels. Data: GROBID's citation corpus (Apache-2.0) with its evaluation files and every reference that also appears in them removed, plus Crossref records (CC0) written out in 11 citation styles with typing noise. Code: github.com/un-seen/citeparse.

Limits

Trained on English-language styles. It labels one reference at a time, so split a bibliography into lines first. "et al." comes out as a literal author.

In the browser

onnx/model.onnx runs in transformers.js. Its logits match PyTorch to within 1.5e-5. A demo is at spaces/hedwigai/citeparse.

Downloads last month
52
Safetensors
Model size
32M params
Tensor type
F32
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for hedwigai/citeparse-32m

Quantized
(17)
this model

Dataset used to train hedwigai/citeparse-32m

Space using hedwigai/citeparse-32m 1

Evaluation results