A public benchmark for Indian UPI fraud detection across 5 scam categories, including adversarial paraphrases and multi-turn manipulation sequences.
Total scenarios: 175 (all split="test")
Scam scenarios: 144 (5 categories × 23–37 each, including 34 novel post-2024)
Non-scam scenarios: 31 (15 benign + 16 borderline for false-positive discipline)
Multi-turn scenarios: 15 (2–4 turn trust-building → info-ask → money escalations)
Adversarial paraphrases: 10… See the full description on the dataset page:
https://huggingface.co/datasets/ujjwalpardeshi/chakravyuh-bench-v0.
