---
license: apache-2.0
arxiv: 2601.06498
---


## Summary
This dataset contains spectral classification tasks built for tool-using models.
Each example includes a full-spectrum image, a chat-style prompt, and ground-truth
labels for binary decisions. The metadata also provides the underlying spectral
arrays (wavelength and flux) to support visualization or custom feature extraction.

## Data Files
All files are under `benchmark/data` and share the same schema.
File suffixes correspond to target classes:

- `cc` -> Carbon Star (C-type) (data_source: `carbon`)
- `cv` -> Cataclysmic Variable (CV)
- `gm` -> M-type Giant (gM)
- `ss` -> S-type Star
- `wd` -> White Dwarf (WD)

## Splits and Sizes

| split | file | rows |
| --- | --- | --- |
| train | train_cc.parquet | 671 |
| train | train_cv.parquet | 711 |
| train | train_gm.parquet | 658 |
| train | train_ss.parquet | 706 |
| train | train_wd.parquet | 618 |
| test | test_cc.parquet | 1321 |
| test | test_cv.parquet | 1458 |
| test | test_gm.parquet | 1321 |
| test | test_ss.parquet | 1442 |
| test | test_wd.parquet | 1212 |

Totals: train 3364, test 6754.

## Data Structure
Each example contains the following fields:

- `images`: list with a single image object
- `data_source`: class key string, e.g. `carbon`, `cv`, `gm`, `ss`, `wd`
- `prompt`: list of chat messages
  - `role`: message role (`system`, `user`, etc.)
  - `content`: message text (the user message includes an `<image>` placeholder)
- `reward_model`: struct with ground-truth labels
  - `ground_truth`: string label (typically `YES` or `NO`)
  - `type`: label source (e.g. `rule`)
- `agent_name`: string identifier for the agent configuration
- `extra_info`: metadata
  - `index`: unique id string
  - `name`: spectrum file name
  - `tools_kwargs.spectral_visualization_tool.create_kwargs`:
    - `flux`: list of floats (length ~3909)
    - `wavelength`: list of floats (length ~3909)
    - `redshift`: float

## Notes
- `images` contains the full-spectrum overview image for the example.
- `extra_info` provides raw arrays to recreate spectrum plots.

## Citation

```bibtex
@misc{Jia2026SpecO3,
  author       = {Minghui Jia and Qichao Zhang and Ali Luo and Linjing Li and Shuo Ye and Hailing Lu and Wen Hou and Dongbin Zhao},
  title        = {Spec-o3: A Tool-Augmented Vision-Language Agent for Rare Celestial Object Candidate Vetting via Automated Spectral Inspection},
  eprint       = {2601.06498},
  archivePrefix= {arXiv},
  primaryClass = {cs.CL},
  year         = {2026},
  url          = {https://arxiv.org/abs/2601.06498},
  doi          = {10.48550/arXiv.2601.06498}
}
```