viktoroo/jobseek-postings-labelled · Datasets

viktoroo/jobseek-postings-labelled · Datasets at Hugging Face jobseek-postings-labelled Gold-standard labelled job postings sampled daily from public company career pages. Produced by a Claude-Code-orchestrated pipeline with specialised Sonnet subagents for section splitting and per-section extraction. The dataset is the substrate for training an improved structured-information extractor for jseek.co. Current row counts by date: 2026-05-09: 8 · 2026-05-08: 8 · 2026-05-07: 8 · 2026-05-06: 7 · 2026-05-05: 8 · 2026-05-04: 7 · 2026-05-03: 8 · 2026-05-02: 7 · 2026-05-01: 7 · 2026-04-30: 7 · 2026-04-29: 8 · 2026-04-28: 8 · 2026-04-27: 8 · 2026-04-26: 8 · 2026-04-25: 14 · 2026-04-24: 28. - Sourcing pipeline: https://github.com/colophon-group/jobseek - Public consumer: https://jseek.co - Routine spec:`docs/15-data-sampling-routine.md` in the source repo. - Source schemas:`apps/crawler/src/labeller/schemas/` in the source repo, mirrored under`schemas/` in this HF repo. Quick start ``` from datasets import load_dataset ds = load_dataset("viktoroo/jobseek-postings-labelled", split="train") row = ds0 print(row"input""title_raw") print(row"labels""globals""profession") ``` Structure ``` data/<YYYY-MM-DD>.jsonl one JSONL file per run date schemas/posting.schema.json top-level record schema schemas/sections.schema.json schemas/section_extract/.schema.json schemas/globals.schema.json schemas/qa.schema.json README.md ``` Only postings with`labelling_meta.qa_verdict == "accepted"` are uploaded. Rejected / in-progress postings stay on the collection machine and are not published. Each row — top-level keys - `id`— UUID of the posting (stable across runs) - `schema_version`— always`1` in this release - `sampled_at`/`labelled_at`— UTC timestamps - `source`— company slug + name, board slug, crawler monitor, source URL (+ host), first-seen timestamp - `input`— verbatim`title_raw`; raw + normalised HTML; plaintext; detected locale; char count; numbe