kamaalg/azerbaijani-instructions
Azerbaijani Instruction Dataset (v0) Azerbaijani (instruction, response) pairs for supervised fine-tuning (SFT) of Azerbaijani language models — part of an open Azerbaijani LLM stack. Instruction data is genuinely scarce for Azerbaijani; this is both training data for our models and a reusable standalone artifact for anyone building Azerbaijani instruction-following models. Contents seeds_az.jsonl — 45 hand-authored, high-quality seed pairs spanning 16 task… See the full description on the dataset page: https://huggingface.co/datasets/kamaalg/azerbaijani-instructions.
Azerbaijani Instruction Dataset (v0)
Azerbaijani (instruction, response) pairs for supervised fine-tuning (SFT) of Azerbaijani language models — part of an open Azerbaijani LLM stack. Instruction data is genuinely scarce for Azerbaijani; this is both training data for our models and a reusable standalone artifact for anyone building Azerbaijani instruction-following models.
Contents
- `seeds_az.jsonl` — 45 hand-authored, high-quality seed pairs spanning 16 task categories. These bootstrap the self-instruct generation and are usable directly as a tiny SFT set.
- `sft_az.jsonl` (generated) — the distilled set produced by
generate_instructions.pyfrom a strong teacher model. Not checked in by default; reproduce it yourself (see below).
Each record:
{"category": "qa", "instruction": "Fotosintez nədir və niyə vacibdir?", "response": "Fotosintez bitkilərin ..."}Categories: qa, summarize, rewrite, explain, classify, brainstorm, extract, translate, reasoning, coding-explain, email, list, grammar, definition, step-by-step, creative.
How it's produced
Self-instruct data distillation. generate_instructions.py samples seed pairs (spread across categories), asks a strong teacher (Anthropic Claude — default claude-sonnet-5 — or OpenAI GPT) for new, diverse, native-Azerbaijani pairs, then applies QC: length bounds, Latin-Azerbaijani script enforcement (rejects Cyrillic/Perso-Arabic), markdown-fence stripping, echo rejection, and exact + near-duplicate deduplication. Validate any file with python sft/validate.py --file <path>.
pip install anthropic
export ANTHROPIC_API_KEY=...
python sft/generate_instructions.py --n 3000 --backend anthropicIntended use
- SFT of Azerbaijani base LMs into instruction-following assistants (response-only loss; see
sft/). - A reusable open Azerbaijani instruction corpus for the community.
Limitations (honest)
- Teacher-distilled ⇒ the generated split inherits the teacher's biases, style, and any factual errors. It is not human-written (the 45 seeds are; the bulk is not).
- Native-speaker spot-checks are recommended before high-stakes use; QC here is automated and heuristic (script, length, dedup), not a fluency/factuality guarantee.
- Latin-script Modern (North) Azerbaijani only.
- v0: the seed set is small and skews toward general/educational tasks; extend
seeds_az.jsonlfor broader coverage (domain-specific, multi-turn, tool-use, etc.).
License
Apache-2.0 for the seeds and code. Note that text generated by a teacher model is also subject to that provider's usage terms; check them before redistribution.
