ansarzeinulla/Nogai-Russian-SFT-Biblical-v2

Nogai-Russian SFT Biblical Corpus v2

Russian↔Nogai translation instructions (ChatML) from human Bible translations by the Institute for Bible Translation (IBT). This is a clean rebuild of v1 with splits that can be used for evaluation. Built with `build_sft_clean.py` (seed 42).

Splits

How it was built from v1

1. 1.4,310 v1 rows reduce to 650 unique pairs once exact duplicates are removed.
2. 2.25 pairs dropped because their Nogai/Russian length ratio is outside 0.5–2.0. These were paired by position, not verse, and at least some are different verses.
3. 3.Split by pair first, then each pair is written in both directions. The reverse of a validation or test example is therefore never in train. Pair overlap between any two splits: 0. 3b. No shared sentences. The same verse can appear in two differently aligned pairs, so pair-level dedupe is not enough. Any validation or test pair whose Russian or Nogai side also occurs in train (or, for test, in validation) is moved to train: 6 pairs. Measured on the published files: 0 validation or test rows share a sentence with train.
4. 4.Test is clean of pre-training text: no test pair contains a Nogai sentence of 40+ characters that occurs in Nogai-Unified-Corpus-v1. 157 of the 625 kept pairs do contain such sentences; they were put in train/validation only.

Schema

{"messages": [
  {"role": "system", "content": "You are a highly accurate bilingual translator for Russian and Nogai."},
  {"role": "user", "content": "Переведи этот текст на ногайский язык: <Russian text>"},
  {"role": "assistant", "content": "<Nogai text>"}
]}

Limitations

• —Bible text only. Models tuned on it tend to answer modern-domain input with religious vocabulary.
• —625 pairs is small, so report test scores with more than one training seed.
• —Pairs come from proportional chunking. Verse-number alignment would give exact pairs.

License and permissions

Card and formatting: CC BY-NC 4.0. The Bible texts belong to IBT and are included for non-commercial research.

Citation

@dataset{nogai_russian_sft_v2_2026,
  author       = {Zeinulla, Ansar},
  title        = {Nogai-Russian SFT Biblical Corpus v2},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/ansarzeinulla/Nogai-Russian-SFT-Biblical-v2}}
}