Preprint introduces open road-safety audit model for Bangladesh

An arXiv preprint describes an 8-billion-parameter vision-language model and a 21,947-record Bangladesh road-safety audit dataset intended for visual risk assessment. The authors say they calibrated a teacher model against field audits before using its structured outputs to train the smaller model. They report that the student outperformed its 31-billion-parameter teacher and Gemini-2.5-Flash in blind expert evaluation, while improving over a zero-shot baseline.