The competition will task participants to provide critical information necessary for understanding the conditions and causes of the incident, using a structured VQA schema where all answers must be numeric according to the provided legend. We encourage participants to use Vision Language Models to answer these questions, but participants can employ whichever models/architectures they believe will be successful.

The prerequisite tasks are:
- Weather and time context: Answer questions about time of day and weather conditions using the legend codes (A: Q1.a–Q1.b).
- Light condition: Classify the lighting condition at the time of the incident using the legend codes (B: Q2).
- Traffic environment: Identify where the incident took place and what facilities are present (C: Q3.a–Q3.b).
- Road configuration: Determine road type and lane configuration, including ego-lane position (D: Q4.a–Q4.c).
- Incident type and entities: Classify the incident category and identify the primary incident entity (who caused or could have prevented the incident) and secondary incident entity (victim/secondary cause/entity hit), including vehicle type when applicable and behaviors (E: Q5.a, Q5.b, Q5.c, Q5.e, Q5.f, Q5.g, Q5.i, Q5.j).
- Incident prevention measures: Predict how the incident most likely could have been prevented by the primary and secondary entities (F: Q6.a–Q6.b).
- Initial point of impact: If entities are vehicles, classify where the initial impact occurred and which side of the impacted area was hit for primary/secondary (G: Q7.a–Q7.d).
- Traffic control and signage: Identify the traffic control or road sign present (H: Q8).
- Road surface and material: Classify road surface condition and road material (I: Q9.a–Q9.b).


More details on the exact submission format (including column names and ordering) can be found in the “Data” tab and the provided sample submission file.