💳 Medical Billing Assistant

Qwen2.5-3B fine-tuned for medical billing assistant

Part of the Medical AI Fine-tuned Model Suite — 16 specialist models, one per task

TL;DR

Takes a clinical encounter description and returns billing-ready output.

INPUT:  Procedure: Office visit, established patient, 25 minutes, moderate complexity.
OUTPUT: CPT/HCPCS: 99214\nDescription: Office/outpatient visit, established, moderate complexity\nWork RVU: 1.92\nNote: Assign appropriate ICD-10-CM diagnosis code. Verify modifier requirements.


Base model	unsloth/Qwen2.5-3B-Instruct
Method	QLoRA, 4-bit NF4, rank 16
Training data	medical-billing-sft — 17,029 real-world rows
Training compute	NVIDIA A40 (48GB), ~1.1h
License	Apache 2.0

Architecture

                  +-------------------------+
  user prompt --> |  Qwen2.5-3B-Instruct  | --> base weights (frozen, 4-bit NF4)
                  |  + LoRA adapter (r=16)  | --> medical-billing-qwen25-3b
                  +-------------------------+
                              |
                              v
                     structured output
                  (code / JSON / classification)

This repo contains only the LoRA adapter (~60MB), not the full merged weights. Load it on top of the base model as shown below — this keeps the download small and lets you swap adapters on one base model in memory.

Intended use

Automate medical billing workflows for outpatient charge capture.

Direct use

Describe a clinical encounter, get back CPT code, modifier, and RVU.

Downstream use

Feed into a revenue cycle management system as a first-pass coding suggestion before human billing review.

Out of scope

Final claim submission without human review. Modifier rules that depend on payer-specific contracts.

This model is not a substitute for a certified medical professional's judgment. Output should be reviewed by a qualified person before being used in a clinical or billing decision.

Quickstart

Option A — Transformers + PEFT

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

base_model = "unsloth/Qwen2.5-3B-Instruct"
adapter    = "AmareshHebbar/medical-billing-qwen25-3b"

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)

messages = [
    {"role": "system", "content": "You are a medical billing assistant. Given a clinical encounter description, return the CPT code, ICD-10-CM code, and modifier if applicable."},
    {"role": "user", "content": "Procedure: Office visit, established patient, 25 minutes, moderate complexity."},
]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to(model.device)
outputs = model.generate(inputs, max_new_tokens=128, temperature=0.1, do_sample=True)
print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))

Expected output:

CPT/HCPCS: 99214\nDescription: Office/outpatient visit, established, moderate complexity\nWork RVU: 1.92\nNote: Assign appropriate ICD-10-CM diagnosis code. Verify modifier requirements.

Option B — Unsloth (2x faster load + inference)

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="AmareshHebbar/medical-billing-qwen25-3b",
    max_seq_length=512,
    load_in_4bit=True,
)
FastLanguageModel.for_inference(model)

messages = [
    {"role": "system", "content": "You are a medical billing assistant. Given a clinical encounter description, return the CPT code, ICD-10-CM code, and modifier if applicable."},
    {"role": "user", "content": "Procedure: ECG with interpretation and report, 12-lead."},
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=128, temperature=0.1, do_sample=True)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Option C — vLLM (production serving, OpenAI-compatible)

vllm serve unsloth/Qwen2.5-3B-Instruct \
    --enable-lora \
    --lora-modules medical-billing-qwen25-3b=AmareshHebbar/medical-billing-qwen25-3b \
    --host 0.0.0.0 --port 8000 --dtype bfloat16

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="not-needed")
response = client.chat.completions.create(
    model="medical-billing-qwen25-3b",
    messages=[
        {"role": "system", "content": "You are a medical billing assistant. Given a clinical encounter description, return the CPT code, ICD-10-CM code, and modifier if applicable."},
        {"role": "user", "content": "Procedure: MRI brain with and without contrast."},
    ],
    temperature=0.1,
)
print(response.choices[0].message.content)

Option D — GGUF / llama.cpp (CPU / edge inference)

This repo ships LoRA adapter weights, not a pre-merged GGUF. To run on llama.cpp, merge first:

pip install unsloth
python -c "
from unsloth import FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained('AmareshHebbar/medical-billing-qwen25-3b', load_in_4bit=False)
model.save_pretrained_gguf('medical-billing-qwen25-3b-gguf', tokenizer, quantization_method='q4_k_m')
"

Training details

Data

Trained on 17,029 examples extracted from real CMS Physician Fee Schedule 2026 — PPRRVU nonQPP procedures with RVU data (source). No synthetic or LLM-generated training data — every example pairs real-world input with its authoritative output.

Split	Rows
Train	13,623
Validation	1,703
Test	1,703

Full extraction pipeline documented on the dataset card.

Hyperparameters

Parameter	Value
LoRA rank (r)	16
LoRA alpha	32
LoRA dropout	0
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Quantization	4-bit NF4 (QLoRA)
Max sequence length	512
Optimizer	paged_adamw_8bit
LR schedule	2e-4, cosine
Gradient checkpointing	Unsloth (smart offload)

Training compute


GPU	NVIDIA A40 (48GB)
Cloud provider	RunPod
Training time	~1.1h (incl. eval + hub push)
Tracking	W&B run
CO2 estimate	self-reported, not measured with a carbon tracker — treat as approximate

Fine-tuned with Unsloth for 2x faster training and reduced VRAM, using TRL's SFTTrainer. Full project: wandb.ai/amareshhebbar-/axiomapper.

Bias, risks & limitations

Data recency. Training data reflects a specific snapshot in time (CMS FY2026 / dataset publish date). Codes, rates, and rules referenced may become outdated as source authorities issue updates — always cross-check against the live authoritative source before high-stakes use.

Failure mode. Like any LLM, this model can produce a plausible-sounding but incorrect output, especially on rare, ambiguous, or highly compound real-world cases that fall outside the training distribution. It does not know when it's wrong.

Language. English-language input only (Hindi-medical model excepted, where Hindi system prompts are used but underlying clinical reasoning data is largely English-sourced).

Not a regulated medical device. This model has not been validated, cleared, or approved by any regulatory body (FDA, CDSCO, or equivalent) as a medical device or clinical decision support tool. It is a research/engineering artifact.

Misapplication risk. Do not use this model as the sole basis for a clinical, billing, or compliance decision affecting a real patient or claim. Do not deploy in an emergency triage context without a human-in-the-loop and clear escalation paths.

FAQ

Q: Can I merge the adapter into the base model for faster inference? Yes — use model.merge_and_unload() after loading with PEFT, or use Unsloth's save_pretrained_merged() method.

Q: Why QLoRA instead of full fine-tuning? The base model already has strong language and medical knowledge from pretraining. QLoRA adapts only ~0.5-1% of parameters, which is enough to specialize the output format and domain without the cost or overfitting risk of full fine-tuning.

Q: Can I fine-tune this further on my own data? Yes, this adapter can be used as a starting checkpoint for continued fine-tuning. Note this may require merging first depending on your training framework.

Q: Why is the output format so strict? Each task was trained on a fixed system prompt and consistent output structure. Following the documented system prompt closely (see Quickstart above) gives the most reliable results — deviating from it may produce inconsistent formatting.

Q: Does this model store or transmit my input data? No. Like any open-weight model, all inference happens locally on your own infrastructure (or wherever you deploy it) — nothing is sent back to the model author.

Troubleshooting

Symptom	Likely cause	Fix
`ValueError: padding_token not set`	Base tokenizer has no pad token	Set `tokenizer.pad_token = tokenizer.eos_token` before inference
Garbled / repeated output	Wrong chat template applied	Make sure you use `tokenizer.apply_chat_template`, not a raw string prompt
CUDA OOM on load	Insufficient VRAM	Use `load_in_4bit=True` (already default above) or reduce `max_seq_length`
Adapter loads but ignores fine-tuning	Base model mismatch	Confirm you loaded the exact base listed above — adapters are not portable across different base models or quantizations

Related models in this suite

Model	Task	Size
icd10-coder-qwen25-7b	ICD-10-CM medical coding	7B
snomed-mapper-qwen25-7b	Clinical concept mapping	7B
icd10-to-drg-qwen25-1b	ICD-10 to DRG reimbursement	1.5B
pmjay-classifier-qwen25-3b	India PM-JAY classification	3B

Full suite overview: AmareshHebbar/medical-ai-model-suite

Changelog

Version	Date	Notes
v1.0	2026	Initial release — QLoRA fine-tune on 17,029 real-world rows

Citation

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