omaraboelmaaty/arabic-cv-scoring-dataset

Arabic CV Scoring Dataset

Dataset Summary

This dataset contains ~7,220 synthetically generated Arabic CVs, each paired with a job category, an ATS (Applicant Tracking System) compatibility score, and a suitability score/class label. It was built to train and evaluate the Arabic CV Analyzer — an NLP pipeline that scores, classifies, and generates improvement suggestions for Arabic CVs targeting the Arab job market, where no equivalent ATS-optimization tooling previously existed.

Dataset Creation

• Source CVs: English resumes from the Kaggle snehaanbhawal/resume-dataset (2,484 rows), translated and rewritten into natural Arabic
• Improvement suggestions: informed by the ArabJobs dataset
• Generation pipeline: local LLM (Ollama, qwen2.5:7b) — cloud providers (Gemini, Groq/Llama 4, HF Inference) were evaluated first but hit rate/quota limits during large-scale generation
• Scoring: a rule-based ATS engine (structural checks + domain keyword matching) combined with LLM-assisted suitability scoring
• Balancing: the original class distribution was heavily skewed toward the top tier (~73% "ممتاز"/excellent). A balancing pass generated additional synthetic CVs for underrepresented classes until each class reached the size of the majority class, growing the dataset from 2,484 to ~7,220 rows. No original rows were removed.

Dataset Structure

7,220 rows × 7 columns. Each row represents one Arabic CV:

Field	Type	Description
id	int	Unique row identifier
category	string	Job category, one of 24 English category keys (see below)
original_cv	string	Original English CV text. Populated only for the first 2,484 rows (the source Kaggle CVs); empty for rows added during class balancing, since those were synthetically generated in Arabic directly
arabic_cv	string	The full Arabic CV text (translated/rewritten, or synthetically generated)
suggestions	string	Arabic-language improvement suggestions for the CV, generated by the LLM
ats_score	int	Rule-based ATS compatibility score, range 0–100
suitability_score	int	Suitability score for the given job category, range 42–88

24 job categories (inherited from the source Kaggle resume dataset): ACCOUNTANT, ADVOCATE, AGRICULTURE, APPAREL, ARTS, AUTOMOBILE, AVIATION, BANKING, BPO, BUSINESS-DEVELOPMENT, CHEF, CONSTRUCTION, CONSULTANT, DESIGNER, DIGITAL-MEDIA, ENGINEERING, FINANCE, FITNESS, HEALTHCARE, HR, INFORMATION-TECHNOLOGY, PUBLIC-RELATIONS, SALES, TEACHER.

Note: this file does not include a pre-computed class label column. The 4-tier class (ضعيف / متوسط / جيد / ممتاز) used by the Arabic CV Analyzer model is derived from suitability_score at training/inference time using dynamic percentile thresholds (P25/P50/P75) rather than stored directly in the dataset.

Intended Use

This dataset is intended for:

• Training Arabic NLP models for resume/CV scoring and classification
• Benchmarking Arabic BERT variants (e.g. AraBERT, CAMeL-BERT) on a domain-specific regression + classification task
• Research on ATS systems and resume screening in non-English markets

Limitations & Biases

• Synthetic scores: suitability scores were generated by an LLM, not human recruiters — they may not perfectly reflect real-world hiring judgments and can carry model-specific biases.
• Category coverage: limited to the 24 categories present in the source Kaggle dataset; CVs outside these domains are out of distribution.
• Language: Arabic only. The dataset does not include English or mixed- language CVs.
• Translation artifacts: since source CVs were originally in English and translated/rewritten into Arabic, some phrasing may read as less natural than organically-written Arabic CVs.

License

Released under Apache 2.0, matching the license of the source Kaggle dataset.

Citation / Attribution

If you use this dataset, please link back to the Arabic CV Analyzer project and the original Kaggle snehaanbhawal/resume-dataset.