Why this dataset
Most public conversational fraud datasets are small, concentrated in non-Indian settings, or labelled only at the conversation level. They can support retrospective classification, but they are not designed to study when a system can first identify a scam as it unfolds.
ICFD-31k addresses this gap by supporting both full-conversation classification and streaming evaluation across diverse fraud scenarios.
What is in the release
The dataset contains more than 31,000 scenario-grounded synthetic transcripts spanning ten fraud categories, including financial impersonation, job scams, and other common forms of conversational fraud.
Each transcript includes a final verdict, chunk-level streaming labels, and an explanatory rationale. Human evaluation reports κ = 0.534, indicating moderate inter-annotator agreement.
Models and evaluation
We fine-tune two RoBERTa baselines. M1 evaluates complete conversations, while M2 makes incremental predictions as new chunks arrive.
On the in-domain evaluation, M1 reaches 99.40% F1 and retains 92.97% F1 on held-out fraud types. These results provide reference baselines for the dataset; they should not be interpreted as evidence that real-time conversational fraud detection is solved.
I presented this work at IJCAI–ECAI 2026 in Bremen.