Fake Job Postings Dataset - Replication Archive for "TabuLLM: Feature Extraction from Tabular Text Data using Large Language Models"

This dataset contains 17,880 job postings labeled as fraudulent (4.84%) or legitimate (95.16%). It is the EMSCAD (Employment Scam Aegean Dataset) originally published by Vidros et al. (2017, Future Internet, doi:10.3390/fi9010006) and distributed via Kaggle (shivamb/real-or-fake-fake-jobposting-prediction). This archive is provided for reviewers and readers of the manuscript "TabuLLM: Feature Extraction from Tabular Text Data using Large Language Models" (submitted to the Journal of Statistical Software), as a convenience mirror to support offline replication without requiring a Kaggle account. File: fake_job_postings.csv - 17,880 rows × 18 columns. Columns include 7 free-text fields (title, location, department, company_profile, description, requirements, benefits), 3 binary indicators, 5 categorical features, 1 numeric identifier (job_id), 1 sparse field (salary_range, 84% missing), and 1 binary target (fraudulent). Version 2 adds pre-computed text embeddings for the 7 free-text columns (title, location, department, company_profile, description, requirements, benefits), generated using the OpenAI Embeddings API (May 2026). Each zip contains 7 NumPy .npy files, one per column, in the order listed above. fraud_openai-small_zenodo.zip — model text-embedding-3-small (1536 dimensions); each file shape: (17880, 1536), dtype float64. fraud_openai-large_zenodo.zip — model text-embedding-3-large (3072 dimensions); each file shape: (17880, 3072), dtype float64. File naming convention: fraud_{model}_{dim}d_{column}.npy. Row order matches fake_job_postings.csv exactly. Version 3 adds pre-computed embeddings from two further models (MiniLM and BGE-Large), generated using tabullm.TextColumnTransformer with default settings: columns concatenated with separator || , no column-name labels prepended. The all-columns files embed the 7 text fields in canonical order: title || location || department || company_profile || description || requirements || benefits. fraud_minilm_zenodo.zip — model sentence-transformers/all-MiniLM-L6-v2 (revision fa97f6e7cb1a59073dff9e6b13e2715cf7475ac9), 384 dimensions. Shape: (17880, 384), dtype float64. Contains 19 files: 7 single-column, 1 all-columns, and 11 column-sweep files (5 forward starting from title; 6 backward starting from benefits). Three sweep endpoints identical to base files are omitted from the archive: fwd_title_to_title = _title; fwd_title_to_benefits = all-columns file; back_benefits_to_benefits = _benefits. fraud_bge-large_zenodo.zip — model BAAI/bge-large-en-v1.5 via HuggingFace Inference API, 1024 dimensions. Shape: (17880, 1024), dtype float64. Contains 8 files: 7 single-column plus 1 all-columns. Single-column files: fraud_{model}_{dim}d_{column}.npy. All-columns file: fraud_{model}_{dim}d.npy. Column-sweep files: fraud_{model}_{dim}d_fwd_{start}_to_{end}.npy (columns from {start} to {end} in canonical order) and fraud_{model}_{dim}d_back_{start}_to_{end}.npy (columns from {start} to {end} in reverse-canonical order). Row order matches fake_job_postings.csv exactly in all files. Version 4 adds pre-computed paper-experiment results. The file results.zip contains outputs from the SoftwareX manuscript experiments, organised into three subdirectories: results/t1_preds/ (per-fold model predictions), results/t2_metrics/ (per-fold evaluation metrics), and results/t3_rolled/ (rolled-up summaries used to produce the manuscript figures). These results were generated by the scripts in the paper_experiments/softwarex_revision/ folder of the public TabuLLM repository (https://github.com/asmahani/TabuLLM), where the full experiment code is available. That folder also includes download_results.py, a convenience script that fetches and extracts this archive automatically via the concept DOI. SHA-256: a54215bd4491705740a856fb1ae5469011dffa8e62ca897dad715e909f39a0b2.