TransCity-VLM Dataset
The TransCity-VLM Dataset provides multimodal smart-city data for traffic, energy, mobility, grid operation, urban context understanding, and map-grounded question answering. It supports the training and evaluation of vision-language models for urban prediction, decision support, conversational QA, and reasoning tasks.
The training data are available at this Hugging Face dataset repository.
Dataset Summary
| Split | Rows / Files |
|---|---|
| test JSONL rows | 308,642 |
| large-scale QA rows | 101,736 |
| multi-turn QA rows | 138,558 |
| COT QA rows | 37,713 |
| map screenshot files | 28,055 |
Dataset Components
| Component | Content | Typical use |
|---|---|---|
test/ | Forecasting-style JSONL records for traffic and energy domains across multiple regions and time windows. | Checkpoint evaluation and reproducibility. |
qa_large_scale/ | Single-turn QA records covering traffic, energy, bus mobility, grid operation, site context, relevance, and explanation tasks. | Instruction tuning and QA training. |
qa_multi_turn/ | Multi-turn smart-city dialogues spanning traffic, energy, bus, grid, and urban context tasks. | Dialogue modeling and conversational QA training. |
qa_cot/ | QA records with reasoning traces for prediction, explanation, relevance, bus, grid, traffic, and energy tasks. | Reasoning-oriented training and evaluation. |
map_screenshots/ | Map screenshot images for Beijing, GLA, Glasgow, Grid, Nezaland, and Singapore. | Visual grounding and map-based VQA. |
preview/ | Lightweight examples from the forecasting, QA, reasoning, dialogue, and GLA image components. | Quick inspection in Dataset Preview. |
Coverage
| Aspect | Description |
|---|---|
| Regions and cities | Beijing, GLA, Glasgow, Singapore, Nezaland, Grid, and regional traffic or electricity collections. |
| Time windows | 1h, 3h, 12h, 24h, and 72h windows, depending on the task and domain. |
| Domains | Traffic flow, electricity and energy, bus operations, grid operation, urban site context, and map-grounded visual context. |
| QA tasks | Prediction, explanation, relevance assessment, decision support, multi-turn dialogue, and chain-of-thought reasoning. |
| Modalities | JSONL text and structured records, together with map screenshot images. |
Format
Each JSONL line is a single JSON object. Image files are provided in the map screenshot folders and can be used as visual context for map-grounded tasks.