The problem
A sponsor in consumer banking wants to explore using small, fine-tuned language models (3B to 8B parameter range) over structured transaction streams to catch fraud patterns their existing rules engine misses. The structured data includes merchant category, amount, location, time, and user history. The problem is a two-step challenge: (1) prompt-engineer and fine-tune a small model to flag suspicious patterns from structured sequences, (2) benchmark against their current rules-engine baseline on a held-out test set. Students would get synthetic labeled data mimicking real distribution.
What success looks like
Fine-tuned model F1 at least on par with baseline rules engine on test set. Writeup of what the model catches that rules miss, and vice versa.
Data access
Synthetic labeled transaction dataset provided. Real data not accessible by design.
Constraints
No real customer data. No PII. Synthetic data only.
Reward
$2,500 team prize + sponsor intro