A retrieval-grounded framework for turning a dashcam accident clip into a structured, three-audience emergency-dispatch plan — dispatcher, civilian bystander, and arriving ambulance crew — with every directive grounded in a cited rule from a 19-rule Japanese regulatory corpus (Road Traffic Act, FDMA urgency-triage manual, JPTEC trauma guidelines), rather than free generation. This repository is the dataset release: 500 human-annotated accident clips and the corpus itself.
FDMA 4-tier triage, assigned from observed medical-care need. Non-Emergency dispatches police only; the other three tiers always dispatch an ambulance.
| Type | n |
|---|---|
| Vehicle-to-vehicle | 143 |
| Vehicle-to-object | 91 |
| Vehicle-to-motorcycle | 90 |
| Vehicle-to-pedestrian | 79 |
| Single-vehicle | 66 |
| Vehicle-to-bicycle | 31 |
Environment: 429 sunny / 24 rainy / 47 snowy · 454 day / 46 night.
Each clip is one JSON file under annotations/: scene observation tags, the
corpus chunk_ids that ground the dispatch, and the 3-POV plan itself.
{
"clip_id": "sft-positive-000001-000018-000050",
"observation_tags": { "collision_type": "vehicle_vehicle", "environment": {...}, ... },
"applicable_chunks": ["jp-doukoho-72-1", "..."],
"pov_structured": {
"pov_1": { "incident_summary": "...", "urgency": "non_urgent|less_urgent|urgent|emergent",
"units": [{ "unit": "police", "text": "...", "cite": [...] }] },
"pov_2": { "do": [...], "warn": [...], "dont": [...] },
"pov_3": { "mechanism": {...}, "load_and_go": "...", "check": [...] }
},
"mmau_source": { "video_id": 1, "type_id": 10,
"path_hint": "CAP-DATA/10/000001/images/" }
}See the repository README for the full field reference and the corpus schema.
This dataset does not redistribute video. Clips are derived from
MM-AU (LOTVS-MM-AU)
— each annotation file is named and tagged with MM-AU's own video_id /
type_id numbering (mmau_source field) so it can be matched back to
the original frames by anyone who has downloaded MM-AU directly from its authors.
Each annotation was made watching the whole source video (every frame, lead-up as well as the accident itself) — urgency triage needs the lead-up context.
corpus.jsonl
records each chunk's source_url.If you use this dataset, please cite the paper: arXiv:2607.23132
@article{adhipradhana2026dispatchrag,
title={DispatchRAG: Grounding Emergency Dispatch Decisions in Real-World Protocols from Traffic Accident Video},
author={Adhipradhana, Muhammad Sulthan and Javanmardi, Ehsan and Bao, Naren and Tsukada, Manabu},
journal={arXiv preprint arXiv:2607.23132},
year={2026}
}