Create README.md

README.md

@@ -0,0 +1,59 @@
+This dataset is a custom-built text classification dataset for detecting fraud risk in internship and entry-level job postings. This dataset is created for an embeddings-based classifier that helps students evaluate whether internship and entry-level job postings may be legitimate, suspicious, or fraudulent.
+
+
+## Dataset Summary
+
+The dataset contains job posting texts labeled into three risk categories:
+
+- `legitimate`: postings that look like normal internship or entry-level job advertisements
+- `suspicious`: postings with warning signs such as vague requirements, missing company information, remote-only work, unclear application process, or unrealistic opportunity language
+- `fraudulent`: postings originally labeled as fraudulent in the source dataset
+
+The final dataset contains 900 examples: 300 legitimate, 300 suspicious, and 300 fraudulent job postings.
+
+## Source Data
+
+This dataset was adopted from the Kaggle public dataset: Real / Fake Job Posting Prediction dataset, which is based on the Employment Scam Aegean Dataset (EMSCAD).
+
+Sources:
+
+- Kaggle: `shivamb/real-or-fake-fake-jobposting-prediction`
+- Original dataset: Employment Scam Aegean Dataset (EMSCAD)
+
+## Custom Processing
+
+The original dataset contains job advertisements with a binary `fraudulent` label. I transformed it into a custom internship-focused dataset by:
+
+1. Combining job fields such as title, location, company profile, description, requirements, benefits, employment type, education, industry, and function into one text field.
+2. Filtering toward internship, student, trainee, junior, assistant, graduate, and entry-level roles.
+3. Preserving fraudulent examples so the model learns scam-related language.
+4. Converting the original binary label into three labels: `legitimate`, `suspicious`, and `fraudulent`.
+5. Adding red-flag annotations such as missing company profile, missing requirements, no company logo, no screening questions, remote/telecommuting, money-transfer language, upfront-fee language, and unrealistic easy-money language.
+6. Creating transparent rule-based suspicious variants from legitimate examples to balance the borderline-risk class.
+7. Balancing the final dataset to 300 examples per class.
+
+## Columns
+
+The dataset includes:
+
+- `id`: row identifier
+- `source_job_id`: original job ID or generated suspicious variant ID
+- `title`: job title
+- `location`: job location
+- `text`: combined job posting text used for model training
+- `label`: target class (`legitimate`, `suspicious`, `fraudulent`)
+- `risk_level`: low, medium, or high
+- `red_flags`: semicolon-separated warning indicators
+- `source_dataset`: source dataset name
+- `source_type`: processed public dataset or rule-augmented suspicious variant
+- `original_fraudulent`: original binary fraud label from EMSCAD
+- `internship_related`: whether the posting matched internship/early-career terms
+
+
+## Limitations
+
+This dataset should not be used as a final authority for deciding whether a job is safe. The `suspicious` class is partly rule-based and may not capture all real-world borderline cases. The dataset is useful for experimentation, model comparison, and demo development, but users should still manually verify company identity, application links, contact methods, and payment-related requests.
+
+## References:
+
+Vidros, S., Kolias, C., Kambourakis, G., & Akoglu, L. (2017). Automatic detection of online recruitment frauds: Characteristics, methods, and a public dataset. *Future Internet, 9*(1), 6.