ISLAM-PO/arabic-to-code-8-langs-3m

Dataset evaluation: See EVALUATION.md for schema checks, indexing status, and language-specific quality limits. Viewer note: default is a lightweight preview; select full to load the complete corpus.

Current Hub Validation Status

• Repository claim: 3,000,000 records
• Dataset Server indexed rows: 1,239,045
• Dataset Server estimate: 1,995,159

The 3M target figure is a raw-repository claim and is not yet fully verified by the Hub index. Validate the JSONL files before publishing a definitive record count.

Arabic-to-Code Dataset - 8 Languages

Train an Arabic-speaking code model with a raw target of 3,000,000 Arabic instruction-to-code pairs across 8 programming languages and 40 JSONL files (11.55 GB). The current Hub index exposes 1,239,045 rows and estimates 1,995,159 rows; the raw target still requires a full JSONL audit.

1. Contents

• 2. Dataset Summary
• 3. Repository Map
• 4. Languages Table
• 5. Categories Table
• 6. Record Schema
• 7. Loading and Training Usage
• 8. Generation and Reproduction
• 9. Considerations and Limitations
• 10. Contributors
• 11. License
• 12. Citation
• 13. Push to Hugging Face Hub

2. Dataset Summary

Attribute	Value
Dataset name	arabic-to-code-8-langs-3m
Programming languages	8 (Python, JavaScript, Java, C++, HTML/CSS, SQL, PHP, Go)
Files	40 *.jsonl files (8 x 5)
Total records	3,000,000 raw target; 1,239,045 currently indexed; 1,995,159 estimated
Total size	11.55 GB (11,832.2 MB)
Format	JSONL, UTF-8, one record per line
Content language	Arabic instructions + Arabic code comments + complete runnable code
Suggested tasks	Arabic-to-code generation, instruction tuning, code explanation in Arabic, multilingual code LLM
License	CC-BY-4.0
Loading	load_dataset with streaming=True (recommended)

3. Repository Map

arabic-to-code-8-langs-3m/
├── README.md
├── LICENSE
├── CITATION.cff
├── .gitattributes
├── 01_Python/          # 375K records
│   ├── 01_basics.jsonl      # 75,000
│   ├── 02_functions.jsonl   # 75,000
│   ├── 03_oop.jsonl         # 75,000
│   ├── 04_algorithms.jsonl  # 75,000
│   └── 05_projects.jsonl    # 75,000
├── 02_JavaScript/      # 375K - same 5 files
├── 03_Java/            # 375K
├── 04_Cpp/             # 375K
├── 05_HTML_CSS/        # 375K
├── 06_SQL/             # 375K
├── 07_PHP/             # 375K
└── 08_Go/              # 375K

graph TD
    ROOT[arabic-to-code-8-langs-3m/<br/>3M target records - 11.55GB]
    ROOT --> PY[01_Python<br/>375K]
    ROOT --> JS[02_JavaScript<br/>375K]
    ROOT --> JV[03_Java<br/>375K]
    ROOT --> CP[04_Cpp<br/>375K]
    ROOT --> WEB[05_HTML_CSS<br/>375K]
    ROOT --> SQ[06_SQL<br/>375K]
    ROOT --> PH[07_PHP<br/>375K]
    ROOT --> GO[08_Go<br/>375K]
    PY --> F1[basics/functions/oop/algorithms/projects]
    F1 --> REC[JSON record<br/>instruction - code - explanation]

pie title Records by language (375K each)
    "Python" : 375000
    "JavaScript" : 375000
    "Java" : 375000
    "C++" : 375000
    "HTML/CSS" : 375000
    "SQL" : 375000
    "PHP" : 375000
    "Go" : 375000

4. Languages Table (8 folders)

#	Folder	Language	Records	Size	What is inside
1	01_Python	Python	375,000	~1,536 MB	Scripts, functions, OOP classes, algorithms, full apps - all runnable
2	02_JavaScript	JavaScript	375,000	~1,403 MB	Node.js + browser code, functions, classes, DOM examples
3	03_Java	Java	375,000	~1,527 MB	Full Main.java programs, collections, OOP
4	04_Cpp	C++	375,000	~1,533 MB	Full g++-ready programs with STL
5	05_HTML_CSS	HTML/CSS	375,000	~1,437 MB	Complete responsive RTL Arabic pages with JS
6	06_SQL	SQL	375,000	~1,520 MB	CREATE + INSERT + SELECT scripts (Postgres/MySQL)
7	07_PHP	PHP	375,000	~1,432 MB	CLI-ready .php scripts
8	08_Go	Go	375,000	~1,444 MB	go run-ready programs
Total	8 folders	—	3,000,000	11,832.2 MB (11.55 GB)	—

5. Categories Table (5 files per language)

Per language; multiply by 8 for the global total (each file = 75,000 per language = 600,000 globally).

#	File	Records/lang	Global (x8)	Content (10 task families each)
1	01_basics.jsonl	75,000	600,000	Sum, factorial, prime check, multiplication table, average, temperature, vowels, reverse string, circle area, leap year
2	02_functions.jsonl	75,000	600,000	Add, max, filter evens, frequency, merge, email validation, password strength, slugify, age calc, pagination
3	03_oop.jsonl	75,000	600,000	Student, bank account, car, employee, library, animal inheritance, cart, hospital, course, inventory
4	04_algorithms.jsonl	75,000	600,000	Bubble sort, binary search, Fibonacci, BFS, Dijkstra, Quicksort, Mergesort, knapsack, KMP, LCS
5	05_projects.jsonl	75,000	600,000	Task manager, calculator, login system, mini shop, clinic booking, library system, sales dashboard, Arabic chatbot, payroll, portfolio site

6. Record Schema

Field	Type	Example	Description
id	string	python-basics-000001	Unique ID: language-category-number
language	string	python	One of 8 languages
category	string	basics	One of basics/functions/oop/algorithms/projects
instruction	string (Arabic)	اكتب كود بايثون كامل...	Arabic task description with N and variable
code	string	def solve_omar...	Complete runnable code with Arabic comments
explanation	string (Arabic)	هذا الحل بلغة بايثون...	Long Arabic explanation + complexity + tests
difficulty	string	مبتدئ/متوسط/متقدم	Beginner / intermediate / advanced (cycled)

Example (shortened):

{
  "id": "python-basics-000001",
  "language": "python",
  "category": "basics",
  "instruction": "اكتب كود بايثون كامل وقابل للتشغيل يقوم بالمهمة التالية: حساب مضروب عدد...",
  "code": "# مثال 1 ...\ndef solve_omar(...):\n ... \nif __name__ == \"__main__\":\n    main()\n",
  "explanation": "هذا الحل بلغة بايثون يشرح ... التعقيد الزمني ... حالات اختبار ...",
  "difficulty": "متوسط"
}

Python samples were executed during validation (returncode 0).

7. Loading and Training Usage

7.1 Stream the full dataset (recommended)

from datasets import load_dataset
ds = load_dataset("ISLAM-PO/arabic-to-code-8-langs-3m", "full", split="train", streaming=True)
print(next(iter(ds)))
py = ds.filter(lambda x: x["language"] == "python")

7.2 One language only

from datasets import load_dataset
ds_py = load_dataset("json", data_files={"train": "hf://datasets/ISLAM-PO/arabic-to-code-8-langs-3m/01_Python/*.jsonl"}, split="train", streaming=True)

7.3 Instruction-tuning format (for training your Arabic code model)

def to_prompt(row):
    return {
        "prompt": f"المستخدم: {row['instruction']}\nالمساعد:\n",
        "completion": f"{row['code']}\n\n# الشرح:\n{row['explanation']}",
    }
# Use with TRL SFTTrainer or axolotl / unsloth chat template

7.4 Fine-tune sketch (Transformers + TRL)

# pip install transformers datasets trl peft
from datasets import load_dataset
from trl import SFTTrainer
ds = load_dataset("json", data_files={"train": "hf://datasets/ISLAM-PO/arabic-to-code-8-langs-3m/01_Python/01_basics.jsonl"}, split="train")
# map with to_prompt(), then SFTTrainer(..., dataset_text_field="prompt")

8. Generation and Reproduction

Generated by generate_code_dataset.py (task pools x per-language code builders x Arabic padding).

python generate_code_dataset.py --demo                              # 4,000 records, ~11 MB
python generate_code_dataset.py --full --per-language 375000 --total-gb 10

Parameter	Default	Meaning
--per-language	375000	Records per language (x8 = 3M target total)
--total-gb	10	Target size; per-record bytes auto-computed

9. Considerations and Limitations

Topic	Details
Synthetic data	Template-generated. Great for bootstrapping an Arabic code model, but mix with real repos (e.g. The Stack, CodeAlpaca-Arabic) before production.
Correctness	Python samples smoke-tested. Other languages follow the same logic but run your own validator for Java/C++/Go/SQL before release.
Repetition	Tasks cycle over 10 families per category with varied N/vars/UIDs. Deduplicate + add real-world seeds for final eval.
Security	No secrets included. Generated table names contain numeric suffixes only. Never train on private keys.
Size	11.55 GB - use streaming or per-language loads.

10. Contributors

Role	Name	Contribution
Owner & concept	ISLAM-PO	Arabic-code training idea, 8 languages + 10 GB requirement
Generation & docs	Muse Spark (AI Assistant)	generate_code_dataset.py, templates, validation, docs
Code review (open)	Open call	Native devs: verify Java/C++/Go/SQL samples, add idiomatic patterns

11. License

CC-BY-4.0. Commercial use allowed with attribution. See LICENSE.

12. Citation

@dataset{arabic_to_code_2026,
  title   = {Arabic-to-Code Dataset: 8 Languages, 3M target Records},
  author  = {ISLAM-PO and Contributors},
  year    = {2026},
  publisher = {Hugging Face},
  version = {1.0.0},
  url     = {https://huggingface.co/datasets/ISLAM-PO/arabic-to-code-8-langs-3m},
  note    = {3,000,000 records, 40 JSONL files, 11.55 GB, CC-BY-4.0}
}

13. Push to Hugging Face Hub

pip install huggingface_hub datasets
huggingface-cli login
cd arabic-to-code-8-langs-3m
git init; git lfs install; git lfs track "*.jsonl"
huggingface-cli repo create arabic-to-code-8-langs-3m --type dataset --yes
git remote add origin https://huggingface.co/datasets/ISLAM-PO/arabic-to-code-8-langs-3m
git add README.md LICENSE CITATION.cff .gitattributes
git commit -m "docs: code dataset card"; git push origin main
git add 01_Python 02_JavaScript; git commit -m "data: batch 1"; git push origin main
# repeat for remaining languages

---

Last updated: 2026-09-03 | Version: 1.0.0 | Status: complete generation target; verify indexed count before publication / 11.55 GB