README.md

metadata

license: apache-2.0
language:
  - ar
pretty_name: Arabic Instruct Curated
tags:
  - instruction-tuning

Arabic Instruct Curated

Curated Arabic instruction-following dataset for SFT of Qwen2.5 LoRAs (RASHID778/qwen2.5-3b-arabic-lora-v2 and -7b-).

Format: JSON Lines, one record per line:

{"messages": [{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}], "source": "alpaca"}

Composition (17,313 rows)

source	kept	raw	notes
Yasbok/Alpaca_arabic_instruct	6,000	52,002	cap reached; filtered short/repetitive/parroting
AhmedBou/Arabic_instruction_dataset_for_llm_ft (train)	3,966	10,000	dropped ar->en translation tasks (Latin-dominated)
yrrhall/Arabic_Aya14200 (aya_collection-aya_dataset)	7,347	13,960	dedup removed 3,135 near-duplicates

Excluded entirely: Mohamed-Sami/instruction-fine-tuning-arabic-dataset (864K rows, template noise + garbled OCR content).

Filtering

• user >= 10 chars, assistant >= 15 chars
• no 8+ repeated-character runs
• no control characters
• Latin/letter ratio > 0.5 dropped (keeps Arabic-dominant samples)
• exact + normalized near-duplicate removal (Alef/Ya/Taa-Marbuta normalization, sha256)
• shuffled with seed 42