metadata
license: mit
task_categories:
- text-generation
language:
- en
tags:
- bagels
pretty_name: Bagels Corpus
size_categories:
- 10M<n<100M
Bagels Corpus v1
This was the dataset mix that was going to be used to train Bagels-30M (which does not exist yet), but after some internal testing, we found it suboptimal.
But, if you still want to download it, clone this repo:
git clone https://huggingface.co/datasets/fromziro/bagels_corpus
cd bagels_corpus_v1
Next, install the required dependencies using pip:
pip install datasets tqdm zstandard
Then run the download script:
python bagels_corpus.py
Stats
| Dataset | Size | Share | License |
|---|---|---|---|
| FineWeb-Edu | 25.0 GB | 34.6% | ODC-By 1.0 |
| DCLM Baseline 1.0 | 15.0 GB | 20.7% | CC BY 4.0 |
| MGA FineWeb-Edu | 4.0 GB | 5.5% | ODC-By 1.0 |
| Stack v2 Edu Filtered | 4.0 GB | 5.5% | Mixed |
| peS2o Filtered | 4.0 GB | 5.5% | Mixed |
| DOAB Filtered | 4.0 GB | 5.5% | CC BY / CC BY-SA |
| PubMed Filtered | 2.0 GB | 2.8% | CC BY / CC BY-SA / CC0 |
| arXiv Abstracts Filtered | 2.0 GB | 2.8% | CC0 |
| StackExchange Filtered | 2.0 GB | 2.8% | Mixed |
| arXiv Papers | 2.0 GB | 2.8% | CC BY / CC BY-SA / CC0 |
| SmolLM Corpus | 2.0 GB | 2.8% | ODC-By 1.0 |
| FineMath | 2.5 GB | 3.5% | ODC-By 1.0 |
| Wikipedia | 1.0 GB | 1.4% | CC BY-SA 3.0 / GFDL |
| Project Gutenberg Filtered | 1.0 GB | 1.4% | Public Domain |
| USPTO Filtered | 0.6 GB | 0.8% | Public Domain |
| Regulations Filtered | 0.6 GB | 0.8% | Public Domain |
| USGPO Filtered | 0.6 GB | 0.8% | Public Domain |
| Total | 72.3 GB | 100.0% |
License
The dataset script is licensed under the MIT License and may be used freely. However, each dataset source has its own license (see above).