UyghurAI/idirak-uyghur-instructions

IDIRAK Uyghur Instructions

A conversational dataset for developing and evaluating the IDIRAK Uyghur AI assistant. Rows use Hugging Face's messages format and include provenance, category, license, and review status fields.

Important status

This version contains 50 schema and pipeline seed examples across training, validation, and test splits. It is not large enough to produce a strong model. Every included row is marked needs_native_review and must be checked by a native Uyghur speaker before training.

Row format

{
  "id": "idirak-train-0001",
  "messages": [
    {"role": "user", "content": "سالام، سەن كىم؟"},
    {"role": "assistant", "content": "سالام! مەن IDIRAK ياردەمچىسى."}
  ],
  "category": "conversation",
  "source": "idirak_seed_v1",
  "license": "other",
  "review_status": "needs_native_review"
}

Splits

• —train: instruction examples used for supervised fine-tuning
• —validation: tuning and overfitting checks
• —test: final pipeline evaluation; never include these rows in training

Create an additional private, native-reviewed evaluation set before publishing performance claims.

Quality policy

• —Keep all text in Unicode NFC form.
• —Preserve correct Uyghur Arabic orthography; do not apply aggressive Arabic or Persian character replacement.
• —Remove duplicates across every split.
• —Record the real source and license for each imported example.
• —Remove personal data and text without redistribution rights.
• —Human-review synthetic translations and answers.

Validate locally

python validate_dataset.py

Review locally

Run the lightweight review interface, correct the text where needed, and mark each row as accepted or rejected. Accepted rows become native_reviewed.

pip install -r requirements-review.txt
python review_app.py

Related resources

• —IDIRAK model
• —IDIRAK Lab
• —idirak.com