tutur90/

slimpajama6b-50k

A 50,000-row training + 1,000-row evaluation quality-filtered subset of DKYoon/SlimPajama-6B. Source proportions are not reweighted — the natural distribution of the dataset is preserved. Documents are shuffled (file order + within-file) then filtered with keep_doc() (see dataset_builders/filters.py):

Filter Criterion

Length 50–100 000 words

Encoding < 1 % replacement/control characters… See the full description on the dataset page: https://huggingface.co/datasets/tutur90/slimpajama6b-50k.