AmareshHebbar/pmjay-classifier-sft

PM-JAY Health Benefit Package Classifier

Part of the AxisMapper Medical AI Suite — 16 domain-specific SFT datasets for fine-tuning medical LLMs.

Built by AmareshHebbar | Studio Ilios / Humanova Minds

---

What this dataset does

Medical specialty + procedure → PM-JAY HBP code, package name, and rate

Why download this

Automate PM-JAY / Ayushman Bharat claim processing. Map procedures to Health Benefit Packages for pre-authorization and reimbursement. Covers 11,140 procedure-package pairs across all specialties.

Dataset stats

Split	Rows
Train	8,912
Validation	1,114
Test	1,114
Total	11,140

Data format

Every row is a messages list in chat format — compatible with Unsloth, TRL SFTTrainer, LLaMA-Factory, and any OpenAI-style fine-tuning pipeline:

{
  "messages": [
    {"role": "system",    "content": "You are a ..."},
    {"role": "user",      "content": "Specialty: Burns Management
Procedure: Thermal burns — Criteria 1 TBSA less than 20%"},
    {"role": "assistant", "content": "Package Code (HBP 2022): BM001
Procedure Code: BM001A
Package Name: Thermal burns
Scheme: AB PM-JAY / HBP 2022
Note: Pre-authorization required."}
  ]
}

Data source

NHA India — HBP 2022 Package Master + PM RAHAT Tier 1 & 2 + Add-on Procedures → https://pmjay.gov.in/about/packages

All data is extracted from authoritative public sources. No LLM-generated or synthetic content.

Who should use this

Hospitals empanelled under PM-JAY, health insurance companies in India, NHA-aligned health IT vendors, Indian public health researchers.

Quick start

from datasets import load_dataset

ds = load_dataset("AmareshHebbar/pmjay-classifier-sft")
print(ds["train"][0])

Fine-tuning example (Unsloth)

from unsloth import FastLanguageModel
from trl import SFTTrainer
from datasets import load_dataset

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="unsloth/Qwen2.5-3B-Instruct",
    max_seq_length=2048,
    load_in_4bit=True,
)

dataset = load_dataset("AmareshHebbar/pmjay-classifier-sft", split="train")

trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    train_dataset=dataset,
    dataset_text_field="messages",
    max_seq_length=2048,
)
trainer.train()

Related datasets in this collection

Dataset	Task	Rows
icd10-coder-sft	ICD-10-CM coding	74.7k
symptom-diagnoser-sft	Symptom → diagnosis	119k
clinical-summarizer-sft	SOAP summarization	30k
discharge-qa-sft	Discharge summary QA	30k
pmjay-classifier-sft	PM-JAY packages	11.1k
radiology-coder-sft	Radiology coding	25k
medical-ner-sft	Clinical NER	16.7k
hindi-medical-sft	Hindi medical QA	19.7k

Citation

@misc{axiomapper2026,
  author    = {Hebbar, Amaresh},
  title     = {AxisMapper: Medical AI Fine-tuning Dataset Suite},
  year      = {2026},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/collections/AmareshHebbar/axiomapper-medical-ai-suite}
}

---

AxisMapper is an open-source project. Star the repo, open issues, and contribute at GitHub.