datdevsteve/nivrachat-mvp-1.0

Nivra ClinicalBERT Patient Context (v2)

Fine-tuned ClinicalBERT for multi-syndrome classification over serialized patient-context notes.

Base model

• medicalai/ClinicalBERT

Task

• Sequence classification: serialized_text -> syndrome label
• Labels (id2label): see id2label.json

Data

• Source: India-focused synthetic/augmented clinical note corpus
• Columns: serialized_text, patient profile fields, label, source_class
• Patch v2: added 300 hard-negative mild-self-limiting samples (heat/sunlight/sports dehydration) to counter chronic_systemic bias.

Training

• Resume checkpoint: ./nivra_clinicalbert_patient_context_v1
• Hyperparameters:
  • epochs: 2
  • lr: 1e-5
  • batch size: 16 train / 16 eval
  • weight decay: 0.01
  • warmup_ratio: 0.05
  • fp16: true
  • metric_for_best_model: weighted_f1
• Framework: Hugging Face Transformers Trainer

Files

• config.json
• model.safetensors
• tokenizer.json, tokenizer_config.json
• label2id.json, id2label.json
• training_args.bin
• README.md (this file)

Inference example

[AGE] 17
[SEX] male
[LOCATION] Jaipur, Rajasthan
[KNOWN_ALLERGIES] none
[PAST_DISEASES] none
[VACCINATION_RECORDS] BCG, OPV, DPT, Hepatitis B, MMR, COVID-19
[SYMPTOMS] weakness, dizziness, sweating after cricket in strong sunlight

Expected: mild_self_limiting

License

• Follow the base model license and ensure synthetic data usage aligns with your policies.