README.md

# TRACE: TimeSeriesRAG Raw Dataset

This is the **raw dataset** accompanying the paper:

 **[TRACE: Grounding Time Series in Context for Multimodal Embedding and Retrieval (NeurIPS 2025)](https://arxiv.org/abs/2506.09114?)**

Feel free to use this dataset for follow-up research and downstream tasks.

---

## Files

| File | Lines | Description |
|------|-------|-------------|
| `event_report.jsonl` | 4,855 | Weather event reports with narrative text |
| `mmts.jsonl` | 44,565 | Multimodal time series samples from weather stations |

---

## Data Format

### `event_report.jsonl`

Each line is a JSON object representing a weather event report:

```json
{
  "event_id": 1065296,
  "event_type": "Debris Flow",
  "state": "CALIFORNIA",
  "cz_name": "MADERA",
  "begin_date_time": "2023-01-10 21:11:00",
  "end_date_time": "2023-01-10 23:11:00",
  "narrative": "A strong low pressure system moved through central California...",
  "ts_dict_index": [12, 13, 14]
}
```

| Field | Type | Description |
|-------|------|-------------|
| `event_id` | int | Unique event identifier |
| `event_type` | string | Type of weather event (e.g., Debris Flow, Flood, Tornado) |
| `state` | string | U.S. state where the event occurred |
| `cz_name` | string | County/zone name |
| `begin_date_time` | string | Event start time (`YYYY-MM-DD HH:MM:SS`) |
| `end_date_time` | string | Event end time (`YYYY-MM-DD HH:MM:SS`) |
| `narrative` | string | Free-text description of the event |
| `ts_dict_index` | list[int] | Indices into `mmts.jsonl` for associated time series samples |

> **Note:** `ts_dict_index` values are 0-based line indices into `mmts.jsonl`, linking each event report to one or more nearby weather station time series.

---

### `mmts.jsonl`

Each line is a JSON object representing a multimodal time series sample from a weather station:

```json
{
  "id": "0",
  "station_id": "USW00025323",
  "latitude": 59.2428,
  "longitude": -135.5114,
  "temperature": [2.2, 1.7, 1.4, ...],
  "precipitation": [1.0, 0.5, 0.5, ...],
  "relative_humidity": [82.0, 85.0, 88.5, ...],
  "visibility": [12.88, 11.26, 6.44, ...],
  "wind_u": [-1.59, 2.25, -0.54, ...],
  "wind_v": [-0.61, 1.3, 3.05, ...],
  "sky_code": [8, 8, 8, ...],
  "DATE": ["2020-11-24T00:00:00", "2020-11-24T01:00:00", ...],
  "mode": "7day_hourly",
  "location": "HAINES BOROUGH,ALASKA",
  "description": {
    "DATE": "The past week from November 24 to November 30, 2020.",
    "location": "...",
    "temperature": "...",
    "precipitation": "...",
    "relative_humidity": "...",
    "visibility": "...",
    "wind_u": "...",
    "wind_v": "...",
    "sky_code": "...",
    "labels": "[Cold, Rainy, Cloudy, Windy]"
  }
}
```

| Field | Type | Description |
|-------|------|-------------|
| `id` | string | Row index (matches position in file, 0-based) |
| `station_id` | string | NOAA weather station identifier |
| `latitude` / `longitude` | float | Station coordinates |
| `temperature` | list[float] | Temperature readings (°C) |
| `precipitation` | list[float] | Precipitation (mm) |
| `relative_humidity` | list[float] | Relative humidity (%) |
| `visibility` | list[float] | Visibility (km) |
| `wind_u` | list[float] | Eastward wind component (m/s) |
| `wind_v` | list[float] | Northward wind component (m/s) |
| `sky_code` | list[int] | Sky cover code (0–8 oktas) |
| `DATE` | list[string] | ISO 8601 timestamps for each hourly reading |
| `mode` | string | Sampling mode (e.g., `7day_hourly` = 7-day window at hourly resolution) |
| `location` | string | Human-readable station location |
| `description` | dict | Natural language descriptions of each channel plus weather labels |

---

## Linking Events to Time Series

The `ts_dict_index` field in `event_report.jsonl` contains a list of line indices (0-based) pointing to rows in `mmts.jsonl`. These identify the weather station time series samples spatially and temporally associated with each event.

```python
import json

with open("mmts.jsonl") as f:
    mmts = [json.loads(line) for line in f]

with open("event_report.jsonl") as f:
    for line in f:
        event = json.loads(line)
        related_ts = [mmts[i] for i in event["ts_dict_index"]]
```

---

## Preprocessed Dataset

A preprocessed version of this dataset (formatted for model training and evaluation) is available for download: [Google Drive Link](https://drive.google.com/file/d/1hX4D91QbXa0UQlgf6Jnf-1ii96gfp1aY/view?usp=sharing)

---

## Citation

If you use this dataset, please cite:

```bibtex
@article{chen2025trace,
  title={Trace: Grounding time series in context for multimodal embedding and retrieval},
  author={Chen, Jialin and Zhao, Ziyu and Nurbek, Gaukhar and Feng, Aosong and Maatouk, Ali and Tassiulas, Leandros and Gao, Yifeng and Ying, Rex},
  journal={arXiv preprint arXiv:2506.09114},
  year={2025}
}
```