README.md

metadata

language:
  - rw
tags:
  - kinyarwanda
  - text
  - pretraining

wandaa-kin-traindata

Training text for the wandaa-kin-gpt Kinyarwanda model.

• kinya_clean.txt, kinya_merged.txt, train_blocks*.npy : first corpus (about 2.9M tokens, mostly textbook text), tokenized with the 8000-token tokenizer map-boy/wandaa-kin-tokenizer
• v3/clean_docs_v3.jsonl : cleaned documents (12,212 docs; fields: source, text) from kigalitoday, igihe, umuryango, cpt_train, firecrawl and the textbook corpus. Markdown, urls, menus, English/French lines and repeated boilerplate removed; curly apostrophes converted to straight ones
• v3/corpus_v3_train.npy, v3/corpus_v3_val.npy : 129-token blocks (6.6M train tokens, 0.19M validation tokens; validation split by whole document)

Warning: sources include scraped news sites. Check each site's terms before reusing the text beyond research.