idirak-model

IDIRAK Model

UyghurAI/idirak-model is the training and release repository for the IDIRAK Uyghur assistant. The repository currently contains the reproducible IDIRAK v1 QLoRA pipeline. It does not contain trained production weights yet.

Intended capabilities

• Natural Uyghur conversation in the Arabic script
• Uyghur, English, Chinese, and Turkish translation assistance
• Summarization and rewriting
• Grounded answers for the IDIRAK document tool
• Safe, honest handling of uncertain or unsupported questions

IDIRAK v1 approach

1. Compare multilingual base models on a native-reviewed Uyghur evaluation set.
2. Fine-tune the best base model with QLoRA and assistant-only loss.
3. Evaluate the adapter against the unchanged base model.
4. Publish the adapter here only after native-speaker review.
5. Point IDIRAK Lab at this repository after trained weights are available.

The default base model is Qwen/Qwen3-4B. You can override it with --model-id, for example google/gemma-3-4b-it after accepting its license.

Train

Use a Linux machine with an NVIDIA GPU. Four-bit QLoRA is not intended to run on the current macOS development machine.

python -m venv .venv
source .venv/bin/activate
pip install -r requirements-train.txt
hf auth login
python train_qlora.py \
  --model-id Qwen/Qwen3-4B \
  --dataset-id UyghurAI/idirak-uyghur-instructions \
  --output-dir artifacts/idirak-qwen3-4b-lora

Review local results before publishing. To upload the trained adapter:

python train_qlora.py \
  --model-id Qwen/Qwen3-4B \
  --dataset-id UyghurAI/idirak-uyghur-instructions \
  --output-dir artifacts/idirak-qwen3-4b-lora \
  --push-to-hub \
  --hub-model-id UyghurAI/idirak-model

Evaluate base models or an adapter

python evaluate.py \
  --model-id Qwen/Qwen3-4B \
  --dataset-id UyghurAI/idirak-uyghur-instructions \
  --output-file evaluations/qwen3-4b-base.jsonl

python evaluate.py \
  --model-id Qwen/Qwen3-4B \
  --adapter-id UyghurAI/idirak-model \
  --dataset-id UyghurAI/idirak-uyghur-instructions \
  --output-file evaluations/idirak-v1.jsonl

Each prediction is saved with blank human-review fields for Uyghur fluency, correctness, instruction following, safety, and an overall score. Translation examples also receive an aggregate chrF score.

Data quality gate

The initial Hub dataset is a small schema and pipeline seed, not sufficient for production training. Every row marked needs_native_review must be reviewed by a native Uyghur speaker. A serious first release should target at least tens of thousands of licensed, deduplicated, reviewed conversations and maintain a separate hidden evaluation set.

Ecosystem

• IDIRAK Lab
• IDIRAK instructions
• IDIRAK AI Ecosystem
• idirak.com

License and provenance

The final model license depends on the selected base model and the rights for all training sources. Do not publish trained weights until both are documented.