TAXI-PRO: Multimodal Temporal Point Process Dataset

📦 Dataset Access

The complete TAXI-PRO dataset is provided as a single ZIP archive:

# Download complete dataset (6.58 GB)
wget https://huggingface.co/datasets/Iroel266/taxi-pro/resolve/main/Manhattan_2000.zip

# Extract
unzip Manhattan_2000.zip

The ZIP archive contains the full dataset, including:

• 2,000 JSON event sequences
• ~72,000 map image patches
• Full directory structure for multimodal training and evaluation

Repository Contents

This Hugging Face repository currently contains:

• Manhattan_2000.zip — the complete dataset archive
• data/ — JSON sequence files kept in the repository for lightweight preview and inspection
• README.md — dataset documentation

To keep the repository lightweight and easier to browse, the full images_unique/ image tree is not retained in the repository. Users who need the full multimodal dataset should download and extract Manhattan_2000.zip.

Important Note on Repository vs. Full Dataset

The Hugging Face repository is designed for:

• dataset discovery
• documentation
• direct inspection of JSON event sequences

The repository is not intended to host the full expanded image tree for browsing. The authoritative full release is the ZIP archive. If you want to reproduce multimodal experiments that require map images, please download and extract Manhattan_2000.zip first.

Dataset Description

TAXI-PRO is a multimodal temporal point process (TPP) dataset constructed by augmenting the classic NYC Taxi dataset with rich visual and textual information. This dataset is introduced in the ICLR 2026 paper "Long-Range Modeling and Processing of Multimodal Event Sequences" (GitHub).

Dataset Summary

TAXI-PRO extends traditional taxi trip data by incorporating:

• Temporal information: Event timestamps and time intervals
• Event types: Pickup (type 2) and dropoff (type 3) events
• Textual descriptions: Natural language descriptions including landmarks, GPS coordinates, passenger counts, and trip distances
• Visual information: 224×224 pixel map image patches centered at each event location, extracted from OpenStreetMap

Each sequence represents a taxi driver's activity over a period of time, with an average sequence length of 36 events.

Key Statistics

• 2,000 sequences from Manhattan taxi drivers
• ~72,000 unique map images (224×224 pixels) organized by region (lower/midtown/upper Manhattan) and event type (pickup/dropoff)
• Multimodal events: Each event contains timestamp, event type, text description, and corresponding map image path
• Rich metadata: GPS coordinates, passenger counts, trip distances, and landmark information

Full Dataset Structure (inside the ZIP archive)

After extracting Manhattan_2000.zip, you will find:

Manhattan_2000/
├── Manhattan_TPP_Dataset_2000_Final_Clean_Flat/  # JSON sequence files
│   ├── driver_XXX.json (2,000 files)
│   └── ...
└── images_unique/                                # Map images
    ├── lower/
    │   ├── pickup/
    │   └── dropoff/
    ├── midtown/
    │   ├── pickup/
    │   └── dropoff/
    └── upper/
        ├── pickup/
        └── dropoff/

Preview Data in This Repository

The data/ directory in this repository contains JSON event sequences for direct inspection and lightweight preview.

Each JSON file contains a sequence of events such as:

[
  {
    "time": 0.0,
    "int_time": 0,
    "text": "Picked up at Theater District (40.759731, -73.961807), 5 passengers",
    "image_path": "images_unique/midtown/pickup/driver_XXX_event_00_3229_2261_224x224.png",
    "TPP_attribute": {
      "event_type": 2,
      "time_since_start": 0.0,
      "time_since_last_event": 0.0
    }
  }
]

Field Descriptions

• time: Normalized timestamp (in hours from start of sequence)
• int_time: Integer time bucket
• text: Natural language description of the event
• image_path: Relative path to the corresponding map image inside the full dataset archive
• TPP_attribute:
  • event_type: 2 for pickup, 3 for dropoff
  • time_since_start: Time elapsed since the first event in the sequence
  • time_since_last_event: Time interval since the previous event

Usage Example

import json
from pathlib import Path
from PIL import Image

# Load a preview sequence from the repository
with open('data/driver_XXX.json') as f:
    sequence = json.load(f)

# Access first event
event = sequence[0]
print(f"Event: {event['text']}")
print(f"Time: {event['time']}")
print(f"Type: {event['TPP_attribute']['event_type']}")

# For image loading, first download and extract Manhattan_2000.zip
img_path = Path('Manhattan_2000') / event['image_path']
img = Image.open(img_path)
img.show()

Dataset Creation

Source Data

The dataset is based on the NYC Taxi dataset (Whong, 2014), which contains taxi trip records from New York City.

Data Augmentation

For each taxi event (pickup or dropoff):

1. Map images were generated by extracting 224×224 pixel patches from OpenStreetMap tiles centered at the event's GPS coordinates
2. Text descriptions were created incorporating:
   • Landmark names (e.g., "Times Square", "Herald Square")
   • GPS coordinates
   • Passenger counts
   • Trip distances (for dropoff events)
   • Origin and destination information (for dropoff events)

Recommended Splits

• Training: 80% (1,600 sequences)
• Validation: 10% (200 sequences)
• Test: 10% (200 sequences)

Benchmark Results

Performance on TAXI-PRO (from ICLR 2026 paper):

Model	RMSE (↓)	ACC (%) (↑)
NHP	0.4494	75.93
SAHP	0.3526	75.37
THP	0.3736	75.31
Language-TPP	0.3376	75.27
MM-TPP	0.3310	77.56

Citation

If you use this dataset, please cite:

@inproceedings{li2026mmtpp,
  title={Long-Range Modeling and Processing of Multimodal Event Sequences},
  author={Li, Jichu and Zhong, Yilun and Li, Zhiting and Zhou, Feng and Kong, Quyu},
  booktitle={International Conference on Learning Representations (ICLR)},
  year={2026}
}

License

MIT License

Contact

For questions or issues, please contact:

• Yilun Zhong: 2023201953@ruc.rdu.cn

Acknowledgments

This dataset builds upon the NYC Taxi dataset (Whong, 2014) and uses map data from OpenStreetMap.

---

Dataset Repository: https://huggingface.co/datasets/Iroel266/taxi-pro
Paper Code: https://github.com/JichuLi/MM-TPP