Synthetic Dataset and Evaluation Results: Validating AI-Generated Content: Security Challenges and Solutions Inspired by OWASP

Synthetic Dataset and Evaluation Results for LLM Output Validation

This repository contains the synthetic dataset and evaluation results used in the Master's thesis: Validating AI-Generated Content: Security Challenges and Solutions Inspired by OWASP by Kamran Khan, Tampere University, 2025.

Contents

1. Dataset

dataset.csv: Synthetic dataset generated for evaluating LLM outputs

Contains 446 samples across 9 categories

Each sample includes: Context, LLMOutput, TrueLabel

Generated using: Deepseek-V3.2-Exp , Gpt-4


2. Evaluation Results

BaselineResulat.csv: Results from the baseline model using only procedural security and schema/structure validator agent(non-llm)

NoJudgeAgentResult: Results ablation study using semantic, validator and Semantic Agent without Judge audit

EvalauationResult: Results from comple