Arabic Instruction Dataset for Language Model Tuning

CIDAR is a culturally relevant instruction dataset for Arabic containing 10,000 instruction-output pairs. The dataset was created by the arbml organization, with the latest update recorded on July 1, —2025.

Use Cases

• Fine-tuning large language models on Arabic instructions using the 10,000 instruction-output pairs.
• Training models for Arabic grammar correction using the 891 grammar-specific instructions.
• Studying the effect of machine-translated data (from Alpagasus via ChatGPT) on model performance for Arabic tasks.
• Developing culturally adapted AI assistants by leveraging the reviewed, culturally relevant Arabic instructions.

Strengths

• Contains 10,000 total instruction-output samples.
• Includes 891 specific Arabic grammar instructions sourced from 'Ask the teacher'.
• All samples were reviewed by approximately 12 human reviewers.

Limitations

• Primary source is machine-translated data (9,109 samples from Alpagasus via ChatGPT), which may introduce translation artifacts.
• Limited to 10,000 samples, which is a modest size for large-scale model training.
• No information provided on the diversity of instruction types or topics covered.

Provenance

Source

arbml organization on Hugging Face.

Collection Method

Created by selecting and machine-translating ~9,109 samples from the Alpagasus dataset and appending ~891 Arabic grammar instructions from the web.

Time Range

null

Freshness

Last updated on 2025-07-01.

Geography

null