Add frozen-encoder BioMedBERT medical triage model

README.md

@@ -0,0 +1,67 @@
+---
+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.

config.json

@@ -0,0 +1,39 @@
+{
+  "add_cross_attention": false,
+  "architectures": [
+    "BertForSequenceClassification"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "bos_token_id": null,
+  "classifier_dropout": null,
+  "dtype": "float32",
+  "eos_token_id": null,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 768,
+  "id2label": {
+    "0": "self_monitor",
+    "1": "consult_gp",
+    "2": "urgent"
+  },
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "is_decoder": false,
+  "label2id": {
+    "consult_gp": 1,
+    "self_monitor": 0,
+    "urgent": 2
+  },
+  "layer_norm_eps": 1e-12,
+  "max_position_embeddings": 512,
+  "model_type": "bert",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "pad_token_id": 0,
+  "problem_type": "single_label_classification",
+  "tie_word_embeddings": true,
+  "transformers_version": "5.12.0",
+  "type_vocab_size": 2,
+  "use_cache": false,
+  "vocab_size": 30522
+}

label_map.json

@@ -0,0 +1,12 @@
+{
+  "label2id": {
+    "self_monitor": 0,
+    "consult_gp": 1,
+    "urgent": 2
+  },
+  "id2label": {
+    "0": "self_monitor",
+    "1": "consult_gp",
+    "2": "urgent"
+  }
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e8bbe2b168456f991aaa666c181f9cdc1a3e868955a9f9c69421414dbb85e1c0
+size 437961724

tokenizer_config.json

@@ -0,0 +1,15 @@
+{
+  "backend": "tokenizers",
+  "cls_token": "[CLS]",
+  "do_lower_case": true,
+  "is_local": false,
+  "local_files_only": false,
+  "mask_token": "[MASK]",
+  "model_max_length": 128,
+  "pad_token": "[PAD]",
+  "sep_token": "[SEP]",
+  "strip_accents": null,
+  "tokenize_chinese_chars": true,
+  "tokenizer_class": "BertTokenizer",
+  "unk_token": "[UNK]"
+}

training_config.json

@@ -0,0 +1,24 @@
+{
+  "base_model": "microsoft/BiomedNLP-BiomedBERT-base-uncased-abstract-fulltext",
+  "adaptation_method": "Frozen Encoder + Trainable Classification Head",
+  "peft_type": "FROZEN_ENCODER",
+  "trainable_head_modules": [
+    "pooler",
+    "classifier"
+  ],
+  "num_labels": 3,
+  "max_length": 128,
+  "num_train_epochs": 8,
+  "learning_rate": 0.0002,
+  "batch_size_train": 16,
+  "batch_size_eval": 32,
+  "weight_decay": 0.01,
+  "warmup_ratio": 0.1,
+  "fp16": true,
+  "early_stopping_patience": 3,
+  "metric_for_best_model": "f1",
+  "train_size": 490,
+  "validation_size": 105,
+  "test_size": 106,
+  "hf_repo_id": "cristian-untaru/frozen-encoder-biomedbert-medical-triage"
+}