ClaimCourt is a LoRA adapter for Mixtral 8x7B Instruct, fine-tuned to perform calibrated medical-necessity review of Indian health insurance claims (IRDAI 2025 framework). Given a clinical case summary it produces a five-dimension clinical review (cost of care, patient history, facility, clinical appropriateness, utilisation pattern), a determination (MEDICALLY SUBSTANTIATED / REQUIRES VERIFICATION / NOT SUBSTANTIATED), and a calibrated confidence score so low-confidence claims can be routed to a human reviewer.

Trained with Adaptive Data by Adaption for the AutoScientist Challenge (Healthcare category).

Results

Metric	Base (Mixtral 8x7B)	ClaimCourt (adapted)
Win rate on IndiaClaimGuard	13	87
Medical-category win rate (held-out)	40	61

Training: SFT, LoRA r=64 / alpha=128, target modules q/k/v/o, 4 epochs on the Grade-A (quality 9.6, 57.7th percentile) 20k-row IndiaClaimGuard clinical dataset. Eval loss 0.672 -> 0.617, monotonically decreasing (no overfitting).

Usage

python

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

base = "mistralai/Mixtral-8x7B-Instruct-v0.1"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(model, "AniketAsla/claimcourt-medical-lora")

prompt = "Inpatient Medical Record Review\n\nPatient presenting with ..."
msgs = [{"role": "user", "content": prompt}]
inputs = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=420, do_sample=False)
print(tok.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))

Out-of-scope

Trained on synthetic data; not a substitute for a licensed claims adjudicator. Do not use for automated claim denial without human review.

Citation

bibtex

@dataset{aslaliya2026indiaclaimguard,
  title={IndiaClaimGuard: Synthetic Indian Health Insurance Fraud Dataset},
  author={Aslaliya, Aniket},
  year={2026},
  url={https://huggingface.co/AniketAsla/claimcourt-medical-lora}
}

Trained with Adaptive Data by Adaption.