README.md

---
library_name: transformers
pipeline_tag: text-classification
base_model: microsoft/BiomedNLP-BiomedBERT-base-uncased-abstract-fulltext
tags:
- medical-triage
- text-classification
- peft
- frozen-encoder
- linear-probe
- sortmed
---

# Frozen Encoder BioMedBERT Medical Triage

This model is a frozen-encoder sequence classifier for three-class medical
pre-triage classification. The pretrained encoder/backbone is kept frozen and
only the classification head is trained.

## Base model

`microsoft/BiomedNLP-BiomedBERT-base-uncased-abstract-fulltext`

## Task

The classifier predicts one of three triage labels:

- `self_monitor`
- `consult_gp`
- `urgent`

## Adaptation method

- Method: Frozen Encoder + Trainable Classification Head
- Base encoder: frozen
- Trainable: classification head (+ pooler for BERT-style models)
- Max sequence length: 128
- Epochs: 8 (early stopping patience 3), best checkpoint by validation macro F1

## Loading

This is a standard Transformers model and loads with the plain API
(no `trust_remote_code` required):

```python
from transformers import AutoModelForSequenceClassification, AutoTokenizer
tok = AutoTokenizer.from_pretrained("cristian-untaru/frozen-encoder-biomedbert-medical-triage")
model = AutoModelForSequenceClassification.from_pretrained("cristian-untaru/frozen-encoder-biomedbert-medical-triage")
```

## Test metrics

```json
{
  "eval_loss": 0.8381163477897644,
  "eval_accuracy": 0.6038,
  "eval_precision": 0.6201,
  "eval_recall": 0.6058,
  "eval_f1": 0.6107,
  "eval_auc": 0.8076,
  "eval_specificity": 0.7995,
  "eval_iou": 0.4444
}
```

This model is intended for academic experimentation and prototyping, not for
standalone clinical decision-making.