Customized Llama-3-8B-Instruct model optimized for SME credit risk assessment. It extracts discrete behavioral risk features from unstructured corporate data to enable risk-adjusted dynamic pricing in DeFi lending protocols.

`madugula/llama3-sme-risk`

Overview

madugula/llama3-sme-risk is a specialized, local-inference AI model serving as a decentralized behavioral oracle. It is architected to operate within the Reverse Kelly Automated Market Maker (rkAMM) lending framework. The model is designed to process unstructured corporate data (e.g., financial narratives, communication logs) and output deterministic, quantitative risk metrics required for solvency-preserving loan pricing.

Model Details

Base Model: Llama-3-8B-Instruct.
Intended Use: Off-chain credit risk assessment for SME (Small and Medium-sized Enterprise) lending.
Inference Strategy: Deterministic edge-computing; intended to run locally on decentralized infrastructure to eliminate API latency and centralized censorship risks.
Hyperparameters:
temperature: 0.1 (Optimized for deterministic, reproducible financial scoring).
top_p: 0.9.

System Prompt & Output Format

To maintain compatibility with smart contract integration (EVM), the model is strictly constrained to output JSON. It uses the following system instruction:

You are a decentralized credit risk oracle operating within the rkAMM framework. 
Your objective is to extract discrete behavioral risk features from unstructured 
corporate data. Analyze the input text and assess risk. 
You must output ONLY a valid JSON object in this exact format:
{
  "pd_penalty": float,
  "risk_factors": ["list", "of", "factors"],
  "confidence": float
}
Do not include any introductory or concluding text, explanations, or markdown formatting.

Reproducibility

This model is built to ensure total transparency in the credit assessment pipeline. To reproduce the environment:

I. Create the Modelfile:

FROM llama3:8b-instruct
PARAMETER temperature 0.1
PARAMETER top_p 0.9
SYSTEM """You are a decentralized credit risk oracle operating within the rkAMM framework. 
Your objective is to extract discrete behavioral risk features from unstructured 
corporate data. Analyze the input text and assess risk. 
You must output ONLY a valid JSON object in this exact format:
{
  "pd_penalty": float,
  "risk_factors": ["list", "of", "factors"],
  "confidence": float
}
Do not include any introductory or concluding text, explanations, or markdown formatting."""

II. Build and Push:

ollama create madugula/llama3-sme-risk -f Modelfile
ollama push madugula/llama3-sme-risk

Ethical Considerations & Disclaimer

This model provides automated credit risk features based on behavioral patterns. It is intended for use as a data input for algorithmic lending protocols (e.g., the rkAMM) and should not be used as the sole basis for legal or final financial lending decisions without human oversight or comprehensive regulatory compliance checks. The PD (Probability of Default) scores are probabilistic estimates and do not guarantee the repayment of assets.

References

Madugula, S. S., De La Rosa, P.E., & Shankar, D. (2026). Dynamic Interest Rate Discovery in Decentralized Finance: A Reverse Kelly Automated Market Maker for Risk-Adjusted Lending.