PrionNER Release Package

This repository accompanies the paper "PrionNER: A Named Entity Recognition Dataset for Prion Disease Biomedical Literature" and provides the public release of PrionNER, a named entity recognition dataset for prion disease biomedical literature. The release includes canonical train/test splits, fine-grained and coarse-grained annotations, synchronized BRAT/JSON/CoNLL formats, and baseline data-processing and evaluation scripts.

The Hugging Face dataset release is available at https://huggingface.co/datasets/dtan/PrionNER.

The package contains:

• the canonical train and test splits
• two annotation granularities: fine and coarse
• multiple synchronized dataset views: raw BRAT, text-only, JSON, and CoNLL
• minimal code for conversion, evaluation, and baseline training
• metadata and supporting documentation

Package At A Glance

• Documents: 317 total
• Train split: 247 documents
• Test split: 70 documents
• Fine-grained schema: 31 defined labels, 30 observed in the released data
• Coarse-grained schema: 15 labels
• Fine/coarse entity count: 4,655 train and 1,650 test
• Discontinuous entities: 97 train and 34 test

The fine-grained schema defines VPSPr, but that label does not appear in the current released splits.

Directory Structure

final_submission_PrionNER/
├── data/
│   ├── raw_text/
│   │   ├── train/              # text-only copies, one .txt per document
│   │   └── test/
│   ├── fine/
│   │   ├── brat/               # fine-grained BRAT annotations
│   │   │   ├── train/
│   │   │   └── test/
│   │   ├── json/               # one JSON file per document
│   │   │   ├── train/
│   │   │   └── test/
│   │   └── conll/              # token-level BIO export
│   │       ├── train.conll
│   │       └── test.conll
│   └── coarse/
│       ├── brat/               # coarse-grained BRAT annotations
│       │   ├── train/
│       │   └── test/
│       ├── json/               # one JSON file per document
│       │   ├── train/
│       │   └── test/
│       └── conll/              # token-level BIO export
│           ├── train.conll
│           └── test.conll
├── metadata/                   # schema files, BRAT config, machine-readable summary
├── code/                       # conversion, evaluation, and baseline scripts
├── docs/                       # dataset card, reproducibility notes, figures, notes
├── LICENSE
├── DATA_LICENSE.md
├── THIRD_PARTY_RIGHTS.md
└── README.md

What Each Data Folder Means

data/raw_text/

This contains text-only document files, one .txt per document.

• train/: 247 files
• test/: 70 files

Use this when you need only the document text without BRAT annotations.

data/fine/

This is the fine-grained annotation release. The same split is provided in three synchronized formats:

• brat/: original BRAT annotation pairs with fine-grained labels
• json/: one JSON file per document, including full text and entity spans
• conll/: BIO-tagged token classification export

Use fine/ if you want the most specific label set.

data/coarse/

This mirrors data/fine/, but with labels collapsed into a smaller coarse schema.

• brat/: coarse BRAT annotations
• json/: coarse JSON export
• conll/: coarse BIO-tagged export

Use coarse/ when you want a simpler label space or a lower-granularity modeling target.

File Formats

BRAT files

BRAT files live under data/fine/brat/ and data/coarse/brat/.

• .txt stores the document text
• .ann stores text-bound annotations
• discontinuous spans use BRAT's semicolon-separated offset format, for example 554 565;588 595

Example:

• data/fine/brat/train/prion_0744.ann

JSON files

JSON files live under data/fine/json/ and data/coarse/json/, with one file per document.

Each JSON document has this structure:

• doc_id: document identifier
• text: full document text
• entities: list of entity objects

Each entity object includes:

• id: original BRAT entity ID such as T11
• label: entity type
• start, end: document-level character offsets
• text: entity surface text
• spans: one or more {start, end} span objects
• is_discontinuous: present when the entity spans multiple non-contiguous segments

The spans array is important because the dataset includes discontinuous annotations.

Examples:

• data/fine/json/train/prion_0000.json
• data/fine/json/train/prion_0744.json

CoNLL files

CoNLL files live under:

• data/fine/conll/train.conll
• data/fine/conll/test.conll
• data/coarse/conll/train.conll
• data/coarse/conll/test.conll

Format:

• one token per line
• TOKEN<TAB>TAG
• blank line between text lines/sentences
• BIO tags such as B-Symptom, I-Generic_Prion, and O

This format is convenient for token-classification baselines. The richer BRAT/JSON views are the better source if you need exact discontinuous structure.

Recommended Entry Points

Choose the folder based on your goal:

• model-ready document objects with character offsets: data/fine/json/ or data/coarse/json/
• token-classification baselines: data/fine/conll/ or data/coarse/conll/
• simplest text-only corpus access: data/raw_text/

Metadata

The main metadata files are:

• metadata/dataset_summary.json: machine-readable dataset counts
• metadata/prion_ner_entity_definitions_fine.json: fine-grained label definitions and examples
• metadata/prion_ner_entity_definitions_coarse.json: coarse label definitions and examples
• metadata/annotation.conf: BRAT UI configuration

Code And Docs

• code/: conversion, evaluation, export, and training scripts
• docs/dataset_card.md: compact dataset overview
• docs/reproducibility.md: commands for rebuilding exports and running baselines
• docs/release_note_licensing.md: short public-facing note explaining the split licensing policy
• docs/figures: examples of discontinuous, nested, and overlapping structures

Quick Start

From this directory:

python3 code/brat_to_json_conll.py --input-dir data/fine/brat/train --formats both
python3 code/evaluate_conll_predictions.py \
  --gold-file data/fine/conll/test.conll \
  --pred-file YOUR_PREDICTIONS.conll \
  --output-dir eval_out \
  --model-name your_model

Citation

If you use PrionNER, please cite the dataset paper:

@misc{dao2026prionnernamedentityrecognition,
      title={PrionNER: A Named Entity Recognition Dataset for Prion Disease Biomedical Literature},
      author={An Dao and Nhan Ly and Thao Tran and Yuji Matsumoto and Akiko Aizawa},
      year={2026},
      eprint={2605.28375},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2605.28375},
}

Licensing and Redistribution

This release uses a split licensing policy.

• Code in code/ is licensed under the MIT License.
• Project-authored documentation, metadata, annotations, and derived structured outputs are intended to be released under CC BY 4.0, to the extent the maintainers hold the necessary rights.
• Underlying article or abstract text will remain in this public release, but it is not blanket-relicensed by the project. See THIRD_PARTY_RIGHTS.md.

If you want the safest reuse path, prefer using the annotation layer, schema files, statistics, and project-authored metadata rather than redistributing the underlying source text.

Release Notes

• Trained weights, large experiment artifacts, caches, and internal workspaces are intentionally excluded from this package.
• The source-text notice in THIRD_PARTY_RIGHTS.md should be preserved in any redistributed copy of this package.