sciencialab/grobid-superconductors-models

grobid-superconductors models

Trained sequence-labelling models used by grobid-superconductors to extract superconductor materials, critical temperatures (Tc), measurement methods, applied pressures and their relations from scientific literature. The models are loaded at runtime through DeLFT (Deep Learning For Text) inside the GROBID pipeline.

This repository hosts the deep-learning model weights (too large for the application repository); the corresponding Wapiti/CRF models and configuration live in the grobid-superconductors repository under resources/models.

Models

Directory	Architecture	Transformer	Task
superconductors-BERT_CRF	BERT + CRF	scibert_scivocab_cased	Superconductor entities (<material>, <tc>, <tcValue>, <pressure>, <me_method>, <class>)
superconductors-mattpuscibert-BERT_CRF	BERT + CRF	MatTPUSciBERT	Same labels as above, materials-science-adapted transformer
material-BERT_CRF	BERT + CRF	scibert_scivocab_cased	Material-name structure (<formula>, <name>, <doping>, <shape>, <substrate>, <fabrication>, <value>, <variable>)
superconductors-BidLSTM_CRF-with_ELMo	BiLSTM + CRF (+ ELMo)	—	Legacy RNN variant (dropped from the default Docker image)

By default grobid-superconductors uses superconductors-BERT_CRF and material-BERT_CRF; superconductors-mattpuscibert-BERT_CRF is selectable at build time via --build-arg TRANSFORMERS_MODEL=mattpuscibert.

Model directory layout (DeLFT ≥ 0.4)

Each DeLFT model directory contains:

• config.json — model configuration and label set
• preprocessor.json — label/feature vocabularies
• model_weights.hdf5 — trained weights
• transformer-config.json + transformer-tokenizer/ — the base transformer configuration and tokenizer (for BERT_CRF models)

Compatibility note. The BERT_CRF directories were updated to the DeLFT ≥ 0.4 on-disk layout (added preprocessor.json, transformer-config.json, transformer-tokenizer/). The trained weights are unchanged — no retraining — and are loaded as-is by DeLFT 0.4.x (used by GROBID 0.9.x). Older DeLFT (0.3.x) is no longer required.

Usage

These models are not meant to be used standalone; they are consumed by the GROBID service. See the grobid-superconductors documentation for how to run the service and configure the model engine (delft vs wapiti) per model.

Citation

If you use these models, please cite grobid-superconductors and, when using the materials-science transformer, MatTPUSciBERT.