# CPT mixing recipe — generated by `python -m src.training.cpt_mix`.
# Weight = target share x effective budget / raw tokens of that source.
# Sample each source in proportion to its weight; the text is not duplicated on disk.
effective_token_budget: 250000000
effective_tokens_estimated: 248965810
raw_tokens: 500965581
sources:
  quran_tafsir:
    raw_tokens: 283615496
    weight: 0.2644
    target_share: 0.3
  hadith:
    raw_tokens: 71585015
    weight: 0.8731
    target_share: 0.25
  quran_meal:
    raw_tokens: 50726328
    weight: 0.9857
    target_share: 0.2
  risale:
    raw_tokens: 49826379
    weight: 0.4014
    target_share: 0.08
  quran_parallel:
    raw_tokens: 43088869
    weight: 0.2321
    target_share: 0.04
  quran:
    raw_tokens: 1435237
    weight: 17.4187
    target_share: 0.1
  quran_morphology:
    raw_tokens: 688257
    weight: 10.8971
    target_share: 0.03