Datasets:

HarithSami
/

arabic-youtube-scripts-sft

language:
  - ar
license: apache-2.0
task_categories:
  - text-generation
tags:
  - arabic
  - youtube
  - scriptwriting
  - sft
  - dialects
  - qwen2
size_categories:
  - n<1K

Arabic YouTube Scripts SFT Dataset (v3 — 126 Multi-Dialect Long-Form Scripts)

A curated, multi-agent rewritten Supervised Fine-Tuning (SFT) dataset of 126 long-form Arabic YouTube scripts (2,500–4,500 characters each) across 4 major Arabic dialects:

خليجية (Gulf): 43 scripts
مصرية (Egyptian): 29 scripts
عراقية (Iraqi): 27 scripts
شامية (Levantine): 27 scripts

🔗 Project Ecosystem (Hugging Face & Kaggle Cross-Links)

Hugging Face Models Trained on This Dataset

GGUF (Q4_K_M for LM Studio / Ollama): HarithSami/qwen2.5-7b-instruct-arabic-yt-gguf
16-Bit Merged Weights (vLLM / Transformers): HarithSami/qwen2.5-7b-instruct-arabic-yt-merged
LoRA Adapter: HarithSami/qwen2.5-7b-instruct-arabic-yt-lora

Kaggle Mirror & Notebooks

Kaggle Dataset: harithsami/arabic-youtube-scripts-sft
Kaggle Training & Export Notebook: harithsami/yt-arabic-scripts-qwen-finetune
Kaggle 4-Dialect Automated Evaluation Suite: harithsami/inference-test

📐 Dataset Schema & XML Structure

Each row in arabic_youtube_scripts_sft_cleaned.jsonl contains:

row_index: Integer (0 to 125)
metadata: {"channel": "...", "title": "...", "dialect": "Gulf | Levantine | Egyptian | Iraqi", "topic": "..."}
messages: Standard ChatML array (system, user, assistant) in 100% Arabic.

Every assistant script strictly follows the 8-part retention hierarchy:

<hook>...</hook>
<intro>...</intro>
<backstory>...</backstory>
<stakes>...</stakes>
<escalation>...</escalation>
<cliffhanger>...</cliffhanger>
<climax>...</climax>
<cta>...</cta>