A Curated Dataset for Saturation-Guided Method for Operationalizing Data Sufficiency in AI Skill Analysis from Job Advertisements

This dataset accompanies the paper "A Saturation-Guided Stopping Rule for Dictionary-Based AI Skill Extraction from Job Advertisements: A Proof-of-Concept Study on Indonesian Job Postings" (PeerJ Computer Science). It contains a curated collection of 2,847 job postings related to Artificial Intelligence (AI) and Machine Learning (ML) roles, collected from JobStreet and LinkedIn between August and December 2025, with the search restricted to Indonesian postings at the query stage. Through a systematic seven-stage filtering pipeline the collection was refined to 51 analysis-ready postings with complete skill or requirement sections, yielding 840 sentence-level units. The filtering prioritises thematic density over representativeness. The final corpus is therefore not a general sample of either platform's posting population; it is a high-density subset in which 58.5% of retained units contain at least one identified skill mention, which is the property the saturation analysis requires. The dataset supports research in computational workforce analytics, skills gap assessment, natural language processing of job descriptions, and labour market informatics for the Indonesian AI sector. Contents Filtering funnel workbook, nine sheets: Sheet Description Records Pipeline Summary Overview of filtering stages and retention rates 7 All Jobs Complete collection with metadata 2,847 Stage 1, Raw Collection Initial unfiltered collection from both platforms 2,847 Stage 2, Deduplication After removing duplicate postings 2,694 Stage 3, IT Relevance Information technology postings only 1,347 Stage 4, AI Relevance Explicit AI or ML skill mentions 248 Stage 5, Content Quality Sufficient description length and structure 82 Stage 6, Language Processable language (English or Indonesian) 75 Stage 7, Final Sample Complete skill section, analysis-ready 51 Sentence-level corpus, three CSV files: File Description Records job_units_clean.csv JobStreet units, 20 postings 334 linkedin_units_raw.csv LinkedIn units as originally assembled 507 linkedin_units_analysed.csv LinkedIn units after the version 2 corrections 506 Each CSV has the fields unit_id, job_id, platform, section, and text. Version 2 This version corrects the record deposited with the original submission. One posting that did not meet the study's own Stage 7 inclusion criterion has been removed, and nine text fields containing assembly artefacts have been repaired. Both the raw and the corrected LinkedIn files are included so that the corrections are auditable. See CHANGELOG.md for details and for the effect on the reported results.