---
license: mit
task_categories:
  - text-generation
language:
  - en
tags:
  - space
  - astronomy
  - astrophysics
  - cosmology
  - training-data
  - llm-training
size_categories:
  - 1B<n<10B
---

# Space LLM Training Data (~1.27 Billion Tokens)

A curated dataset of space and astronomy text for training language models, containing approximately **1.27 billion tokens** collected from academic papers, arXiv abstracts, and educational web content.

## Dataset Summary

| File | Size | Est. Tokens | Source |
|------|------|-------------|--------|
| `jsalt_astroph_full.txt` | 2.88 GB | ~862M | 271K full astrophysics papers (abstract + introduction + conclusions) |
| `arxiv_astro_full.txt` | 360 MB | ~108M | 284K arXiv paper abstracts (8 space categories) |
| `fineweb_space_combined.txt` | 1.02 GB | ~305M | 95K space-related educational web texts |
| **Total** | **4.16 GB** | **~1.27B** | |

## Sources

### 1. JSALT Astrophysics Dataset (`jsalt_astroph_full.txt`)
- **Source**: [charlieoneill/jsalt-astroph-dataset](https://huggingface.co/datasets/charlieoneill/jsalt-astroph-dataset)
- **Content**: Full paper text (abstract, introduction, conclusions) from 271,544 astrophysics papers
- **Categories**: astro-ph (all subcategories)
- **Format**: Documents separated by `---`

### 2. arXiv Space Categories (`arxiv_astro_full.txt`)
- **Source**: [permutans/arxiv-papers-by-subject](https://huggingface.co/datasets/permutans/arxiv-papers-by-subject)
- **Content**: Title + abstract from 283,699 papers across 8 space-related categories:
  - `astro-ph.CO` (Cosmology & Nongalactic Astrophysics)
  - `astro-ph.GA` (Galaxy Astrophysics)
  - `astro-ph.HE` (High Energy Astrophysical Phenomena)
  - `astro-ph.IM` (Instrumentation & Methods)
  - `astro-ph.SR` (Solar & Stellar Astrophysics)
  - `astro-ph.EP` (Earth & Planetary Astrophysics)
  - `gr-qc` (General Relativity & Quantum Cosmology)
  - `physics.space-ph` (Space Physics)
- **Format**: Documents separated by `---`

### 3. FineWeb-Edu Space Filtered (`fineweb_space_combined.txt`)
- **Source**: [HuggingFaceFW/fineweb-edu](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu)
- **Content**: ~95K educational web texts filtered for space/astronomy keywords
- **Filtering keywords**: astronomy, astrophysics, black hole, galaxy, cosmology, planet, solar system, dark matter, supernova, nasa, orbit, telescope, universe, big bang, stellar, exoplanet, space, rocket, spacecraft, moon, satellite, nebula, quasar, etc.
- **Format**: Documents separated by `---`

## Intended Use

This dataset is designed for training and fine-tuning language models focused on space, astronomy, and astrophysics domains. It was specifically curated for the [Space LLM](https://huggingface.co/Ashu9675/space-llm-10m) project.

## Token Count Estimation

Token counts are estimated using the heuristic of ~3.5 characters per token for mixed academic and web text.

## How to Use

```python
from datasets import load_dataset

# Load from HuggingFace
dataset = load_dataset("Ashu9675/space-llm-training-data")

# Or read the text files directly
with open("jsalt_astroph_full.txt") as f:
    documents = f.read().split("\n---\n")
```

## License

- **arXiv content**: MIT (as per source dataset licenses)
- **FineWeb-Edu content**: ODC-BY (as per FineWeb-Edu license)
- **Combined dataset**: MIT

## Limitations

- Token counts are estimates, not exact
- Some documents may contain LaTeX formatting from academic papers
- FineWeb content is keyword-filtered and may include some false positives
- The jsalt dataset includes only abstract, introduction, and conclusions (not full paper body)