Modelpublic

AmareshHebbar/medical-ai-model-suite

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face

0likes

Model Card

Medical AI Fine-tuned Model Suite

A collection of 16 Qwen2.5 models fine-tuned with QLoRA, one per medical/healthcare task — ICD-10 coding, billing, clinical documentation, India's PM-JAY scheme, and more. Built by Amaresh Hebbar.

Collection page: [link to your HF collection here]

Every model uses the same approach: a real public data source (CMS, NHA India, peer-reviewed biomedical corpora — no synthetic or LLM-generated training data), QLoRA fine-tuning on Qwen2.5, and a narrow, well-defined task with a strict system prompt. The goal is a set of small, deployable specialists rather than one large general-purpose medical model.

Why specialist models instead of one big model

A single general medical LLM has to be evaluated, monitored, and trusted across every task it might be asked to do. These models are scoped to one job each — ICD-10 coding, PM-JAY classification, radiology report coding — so each one is small enough to self-host cheaply, easy to evaluate against a clear ground truth, and safe to swap out independently if a better version ships later.

The 16 models

Model	Task	Base	Rows
icd10-coder-qwen25-7b	ICD-10-CM medical coding	Qwen2.5-7B	74,719
snomed-mapper-qwen25-7b	Clinical terminology mapping	Qwen2.5-7B	74,719
clinical-summarizer-qwen25-7b	SOAP note summarization	Qwen2.5-7B	30,000
symptom-diagnoser-qwen25-7b	Symptom → differential diagnosis	Qwen2.5-7B	119,467
discharge-qa-qwen25-3b	Discharge summary Q&A	Qwen2.5-3B	30,000
radiology-coder-qwen25-3b	Radiology report coding	Qwen2.5-3B	25,090
medical-ner-qwen25-3b	Clinical named entity recognition	Qwen2.5-3B	16,671
hindi-medical-qwen25-3b	Hindi medical reasoning	Qwen2.5-3B	19,704
cpt-coder-qwen25-3b	CPT/HCPCS procedure coding	Qwen2.5-3B	17,029
medical-billing-qwen25-3b	Medical billing assistant	Qwen2.5-3B	17,029
pmjay-classifier-qwen25-3b	India PM-JAY package classification	Qwen2.5-3B	11,140
pharmacy-ner-qwen25-1b	Drug entity extraction	Qwen2.5-1.5B	3,500
ayurveda-icd-qwen25-1b	Ayurveda to ICD-10 bridge	Qwen2.5-1.5B	3,002
insurance-classifier-qwen25-1b	Stark Law DHS compliance	Qwen2.5-1.5B	1,601
icd10-to-drg-qwen25-1b	ICD-10 → MS-DRG reimbursement	Qwen2.5-1.5B	5,385
loinc-coder-qwen25-1b	Lab test CPT coding	Qwen2.5-1.5B	2,179

Training method

All 16 models share the same recipe:


Fine-tuning method	QLoRA, 4-bit NF4 quantization, rank 16, alpha 32
Training framework	Unsloth + TRL `SFTTrainer`
Target modules	qproj, kproj, vproj, oproj, gateproj, upproj, down_proj
Hardware	Single NVIDIA A40 (48GB)
Optimizer	pagedadamw8bit, cosine LR schedule
Experiment tracking	Weights & Biases

Model size is matched to dataset size — large datasets (>50k rows) get Qwen2.5-7B, mid-size (10k–50k) get Qwen2.5-3B, and smaller specialist datasets (<10k) get Qwen2.5-1.5B. This keeps inference cost proportional to task complexity instead of running every task through the same large model.

Data sources

Every dataset behind these models is built from real authoritative public data — CMS (ICD-10-CM, MS-DRG, Physician Fee Schedule, HCPCS), NHA India (PM-JAY HBP 2022, PM RAHAT), and peer-reviewed biomedical corpora (chat_doctor, augmented-clinical-notes, drugprot). No synthetic or LLM-generated training data. Full extraction pipelines and column-level provenance are documented on each dataset card.

How to use any model in this suite

Each model is a LoRA adapter on top of its base Qwen2.5 model. Load with PEFT:

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base_model = "unsloth/Qwen2.5-7B-Instruct"  # match the base size for the model you're using
adapter    = "AmareshHebbar/icd10-coder-qwen25-7b"

tokenizer = AutoTokenizer.from_pretrained(base_model)
model = AutoModelForCausalLM.from_pretrained(base_model, torch_dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(model, adapter)

See each model's individual card for its exact system prompt, example input/output, and recommended serving setup (Transformers, Unsloth, or vLLM).

Limitations

These are narrow specialist models, not general medical assistants. Each model only performs the single task it was trained on — using it outside that task will produce unreliable output. None of these models are a substitute for a licensed medical or billing professional; all output should be reviewed by a qualified person before being used in a clinical, billing, or compliance decision.

Citation

bibtex

@misc{medicalai2026,
  author    = {Hebbar, Amaresh},
  title     = {Medical AI Fine-tuning Suite},
  year      = {2026},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/AmareshHebbar}
}

Contact

—GitHub: amareshhebbar
—LinkedIn: gvamaresh
—HuggingFace: AmareshHebbar