metadata
license: mit
Bulgarian BabyLM Dataset curated by Mila Marcheva (University of Cambridge).
28,467,275 tokens (excluding punctuation)
Dataset Overview
A sentence-level corpus drawn from scanned Bulgarian children's text. Each row represents one segmented sentence, its tokenization, the source URL, and its token count.
Data Sourcing
We additionally release information about the source of sentence in the Bulgarian BabyLM Dataset, which you can find here: https://huggingface.co/datasets/milamarcheva/bulgarian_cds_lg.
Data Schema
| Column | Type | Description |
|---|---|---|
| MainSentencised | string | The raw, sentence-segmented text (in Bulgarian). |
| TokenisedSent | list[string] | The sentence split into word-tokens (lowercased, stripped). |
| SourceLink | string (URL) | . |
| UttLen | int | Token count (length of TokenisedSent). |
Best Practices
We recommend Group by SourceLink to preserve document boundaries. However, this dataset has minimal preprocessing to allow contributors to the Multilingual BabyLM Shared Task to determine their own preprocessing practices.
Written by Suchir Salhan & Mila Marcheva