Sampled version of cerebras/SlimPajama-627B.
Since the original data was shuffled before chunking, I only downloaded train/chunk1 (of 10 total) and further sampled 10%. This should result in roughly 6B tokens, hence SlimPajama-6B.
The dataset is 24GBs in storage size when decompressed (original dataset is over 2TBs) and has 5489000 rows.
The validation set and test set were sampled as well.
Data source proportions for SlimPajama-627B and SlimPajama-6B
For sanity purpose, I caluclated the byte proportion of the sampled version.
| Data source | SlimPajama-627B | SlimPajama-6B |
|---|---|---|
| Commoncrawl | 52.2% | 54.1% |
| C4 | 26.7% | 28.7% |
| GitHub | 5.2% | 4.2% |
| Books | 4.2% | 3.7% |
| ArXiv | 4.6% | 3.4% |
| Wikpedia | 3.8% | 3.1% |
| StackExchange | 3.3% | 2.8% |
Please refer to the original dataset for other info.
@misc{cerebras2023slimpajama,
author = {Soboleva, Daria and Al-Khateeb, Faisal and Myers, Robert and Steeves, Jacob R and Hestness, Joel and Dey, Nolan},
title = {{SlimPajama: A 627B token cleaned and deduplicated version of RedPajama}},
month = June,
year = 2023,
howpublished = {\url{https://www.cerebras.net/blog/slimpajama-a-627b-token-cleaned-and-deduplicated-version-of-redpajama}},
url = {https://huggingface.co/datasets/cerebras/SlimPajama-627B},
}