---
license: cc-by-nc-4.0
dataset_info:
  features:
  - name: file_name
    dtype: string
  - name: RA
    dtype: float64
  - name: DEC
    dtype: float64
  - name: TARGETID
    dtype: float64
  - name: Z
    dtype: float64
  - name: SPECTYPE
    dtype: string
  - name: ZWARN
    dtype: float64
  - name: LOG_MSTAR
    dtype: float64
  - name: sSFR
    dtype: string
  - name: Z_MW
    dtype: string
  - name: TAGE_MW
    dtype: string
  - name: AVG_SFR
    dtype: string
  - name: PROVABGS_LOGMSTAR_BF
    dtype: float64
  - name: MAG_G
    dtype: float64
  - name: MAG_R
    dtype: float64
  - name: MAG_Z
    dtype: float64
  - name: PROVABGS_THETA_BF
    dtype: string
  - name: flag_provabgs
    dtype: string
  - name: id_str
    dtype: 'null'
  - name: image
    dtype: 'null'
  - name: smooth-or-featured-dr5
    dtype: 'null'
  - name: disk-edge-on-dr5
    dtype: 'null'
  - name: has-spiral-arms-dr5
    dtype: 'null'
  - name: bar-dr5
    dtype: 'null'
  - name: bulge-size-dr5
    dtype: 'null'
  - name: how-rounded-dr5
    dtype: 'null'
  - name: edge-on-bulge-dr5
    dtype: 'null'
  - name: spiral-winding-dr5
    dtype: 'null'
  - name: spiral-arm-count-dr5
    dtype: 'null'
  - name: merging-dr5
    dtype: 'null'
  - name: flag_galaxy_class
    dtype: string
  - name: Z_astroclip
    dtype: float64
  - name: flag_desi_ls
    dtype: string
  - name: qualitative_physical_caption
    dtype: string
  - name: quantitative_physical_caption
    dtype: string
  - name: semantic_caption
    dtype: string
  - name: qualitative_integrated_caption
    dtype: string
  - name: quantitative_integrated_caption
    dtype: string
  - name: spectrum_category
    dtype: string
  - name: images_category
    dtype: string
  splits:
  - name: train
    num_bytes: 2496393
    num_examples: 10000
  download_size: 453719
  dataset_size: 2496393
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---
# Go to the MOON🌕  
![galaxylm](./teaser.png)    

> [!NOTE]
> The preview shows only a subset of the first 10,000 items.  
> After finishing the project, we will share all the code on GitHub.

# GalaxyCaption🤖
**Blog🔥: [GalaxySVLM-시리즈-1](https://kyujinpy.tistory.com/188)**  
**Github: [GalaxySVLM](https://github.com/Marker-Inc-Korea/GalaxyLM/tree/main)**


## Three types captions
- **Physical_caption (Spectrum ↔ Text)**: Specturm 정보와 galaxy의 physical property (redshift, age, metallicity, ...)가 대응되는 caption을 만들자.    
-> Qualitative: 물리적 property들에 대해서 정성적으로 작성한 caption.    
-> Quantitative: 물리적 property들을 정량적으로 작성한 caption.    
- **Semantic_caption (Image ↔ Text)**: Image 정보를 토대로 galaxy semantic 정보에 대한 caption을 만들자.    
- **Integrated_caption (Image ↔ Text ↔ Spectrum)**: Galaxy의 image 및 spectrum을 통해 semantic 및 physical property를 모두 표현한 caption.    
-> Qualitative: Physical_Qualitative_caption + Semantic_caption  
-> Quantitative: Physical_Quantitative_caption + Semantic_caption  

# Summary

| Type | # of dataset |
|---|---|
| All | 1,531,756 |
| --- | --- |
| naïve Spectrum (for `specencoder` training) | 1,303,697 |
| Image | 374,966 |
| --- | --- |
| Spectrum ↔ Text | 128,149 |
| Image ↔ Text | 245,015 |
| Spectrum ↔ Image | 153,084 |
| Spectrum ↔ Text ↔ Image | 18,776 |
> There is duplicate galaxy cases.  
> Final dataset: [galaxy_integrated_caption_dataset_final.csv](https://huggingface.co/datasets/Markr-AI/GalaxyCaption/tree/main/excel). (`re-labeling`)  

## Be careful
This repo only produces `Excel` files.
If you want to download all of components (e.g., images and spectra), please follow [dataset.md](https://github.com/Marker-Inc-Korea/SpaceLM/tree/main/dataset).

> [!IMPORTANT]
> `GalaxyCaption` 데이터셋의 `image` 및 `spectrum`은 용량이 너무 크기 때문에 따로 제공되지 않습니다.  
> 데이터셋을 직접 다운받고 제작하고 싶으신 분은 [dataset.md](https://github.com/Marker-Inc-Korea/SpaceLM/tree/main/dataset)를 참고해주세요.

# References
- DESI-EDR (spectrum dataset)
- DESI-LS (image-spectra by AstroCLIP)
- GZ-Decals-5 (galaxy classification)
- PROVABGS (galaxy property)

## Future-work
- GalaxyInstruct
- GalaxyBench