Biomedical NER BiLSTM-CRF Workspace

A small, readable biomedical Named Entity Recognition (NER) project for pharmacy/medical research coursework. It trains a custom PyTorch BiLSTM-CRF model on the Hugging Face tner/bc5cdr dataset, which contains Chemical/Drug and Disease entities.

This is intentionally not SOTA. It is designed to be understandable, runnable from a terminal/Termux/Colab, and easy to discuss in a graduation research project.

Quick start

python setup_ner_workspace.py --write-only
python -m pip install -r requirements.txt
python train.py --epochs 3 --batch-size 16 --max-train-samples 4000
python predict.py --checkpoint checkpoints/best_model.pt

Or let the master script write files, install packages, and start training:

python setup_ner_workspace.py --epochs 3 --batch-size 16 --max-train-samples 4000

Termux notes

Install Python and required build tools first if needed:

pkg update
pkg install python clang rust git
python -m pip install --upgrade pip
python -m pip install -r requirements.txt

On low-memory devices, use smaller settings:

python train.py --epochs 2 --batch-size 4 --embedding-dim 64 --hidden-dim 128 --max-train-samples 1000

Colab notes

Upload this folder or clone your repository, then run:

!pip install -r requirements.txt
!python train.py --epochs 5 --batch-size 32
!python predict.py --checkpoint checkpoints/best_model.pt

Files

• setup_ner_workspace.py — one-command master script that writes all project files.
• data.py — downloads tner/bc5cdr, builds vocab/tag maps, pads batches.
• model.py — custom BiLSTM-CRF architecture.
• metrics.py — entity-level precision, recall, and F1.
• train.py — training/evaluation/checkpoint pipeline.
• predict.py — interactive CLI inference.