shylane/healthcare-fraud-detection-grpo

Healthcare Fraud Detection with GRPO

⚠️ PRELIMINARY MODEL - 25 Training Steps Only

This model is a proof-of-concept for healthcare claims fraud detection using Group Relative Policy Optimization (GRPO). It has only been trained for 25 steps and requires additional training (200-500+ steps recommended) for production use.

Model Details

• Base Model: Qwen/Qwen2.5-0.5B-Instruct (CORRECTED from Llama 3.2 1B)
• Training Method: GRPO (Group Relative Policy Optimization)
• Training Steps: 25 (preliminary)
• Reward Improvement: +38% (from -40.74 to -25.35)
• Initial Entropy: 2.27 (high uncertainty)
• Hardware: NVIDIA RTX 4060 8GB
• Quantization: 4-bit with LoRA adapters

Environment: Healthcare Claims Fraud Detection

OpenEnv-compatible RL environment with:

• LLM-Native Actions: Text generation (Decision + Rationale + Evidence + Recommendation)
• Multi-Component Reward: Decision (40%) + Rationale (30%) + Evidence (20%) + Efficiency (10%)
• Hybrid Data: 60% Synthea synthetic + 40% CMS SynPUF real Medicare claims
• 8 Fraud Patterns: Upcoding, unbundling, phantom billing, etc.
• A2A Protocol: Compatible with Green Agents / AgentBeats

Usage

from transformers import pipeline

# Load model
generator = pipeline(
    "text-generation", 
    model="shylane/healthcare-fraud-detection-grpo", 
    device=0
)

# Generate fraud assessment
prompt = '''Assess this healthcare claim for fraud:
Provider: Dr. Smith
Service: MRI Scan
Amount: $2,500
...'''

result = generator(prompt, max_length=300, temperature=0.7)
print(result[0]["generated_text"])

Training Results

Preliminary (25 steps):

• Shows +38% improvement over random baseline
• Demonstrates environment provides meaningful learning signal
• Entropy decreasing (model becoming more confident)

For Production:

• Requires 200-500+ additional training steps
• Consider larger model (1B-3B parameters) for complex reasoning
• Multi-GPU training recommended for faster convergence

Limitations

⚠️ This is NOT a production-ready model:

• Only 25 training steps completed
• High initial entropy (2.27)
• Still operating at loss (negative rewards)
• Small model size (0.5B parameters)
• Limited to synthetic/hybrid data

Challenge Entry

This model is an entry for the OpenEnv Student Challenge 2026 sponsored by Meta PyTorch, Hugging Face, and Unsloth AI.

Innovation: Moving from discrete action spaces to LLM-native text generation with structured reasoning for healthcare fraud detection.

Code & Documentation

• Repository: https://github.com/shylane/healthcare-openenv-challenge
• Blog Post: See docs/BLOG_DRAFT.md in repository

Framework Versions

• PEFT: 0.18.1
• TRL: 0.27.1
• Transformers: 5.0.0
• PyTorch: 2.6.0+cu124

Citation

@misc{healthcare-fraud-grpo,
  title={Healthcare Fraud Detection with GRPO},
  author={OpenEnv Challenge Entry},
  year={2026},
  howpublished={\url{https://huggingface.co/shylane/healthcare-fraud-detection-grpo}}
}