Dataset · GOAL-Lab
LLM fraud propensity benchmark (FraudBench)
GOAL-Lab/FraudBench
159 legally vetted scenarios from real US courts to test models for deception and fraud propensity.
What's inside
An English text benchmark with under 1,000 entries, including 159 scenarios (101 on misrepresentation and 58 on concealment). Data is compiled from 112 real US court cases between 2014 and 2025 and verified by legal experts for three evaluation formats: binary questions, open-ended questions, and agent tasks.
What you can do with it
- Auditing LLM safety and checking propensity to assist in fraud
- Testing AI agents in legally sensitive scenarios
- Tasks
- Question answering, Text generation
- Languages
- English
- Size
- up to 1K rows
- Downloads this month
- 765
- Likes
- 3
- Access
- on request on Hugging Face
How to load it
from datasets import load_dataset
ds = load_dataset("GOAL-Lab/FraudBench")