Abstract:With the rapid growth of recruitment content online, reliable competency extraction in Chinese remains under-resourced. We present Chinese-SkillSpan, an ESCO-aligned, span-level dataset for Chinese job advertisements under a flat SKTL taxonomy (Language/Skill/Knowledge/Transversal). Spans are non-nested and non-overlapping, annotated to minimal-sufficient boundaries, and linked to ESCO concept IDs. A hybrid pipeline combines LLM assistants to draft spans and types with majority-plus-confidence aggregation, followed by double/triple human adjudication in Doccano, yielding Silver and Gold layers with measured agreement. We release a unified scorer supporting Exact and Relaxed span-F1 and concept accuracy, together with fixed IID and industry/time-shifted OOD splits to foster robust evaluation. Baselines cover ESCO lexicon matching, supervised sequence labeling (CRF and span-based models), cross-lingual zero-shot encoders, and an LLM stack with context-anchored prompting and constrained decoding; a context-first fallback priority L > S > K > T resolves borderline cases. Experiments on the released splits highlight gains from title+local-context prompting and reveal challenges on long-tail technical and socio-behavioral skills. Data and code are publicly available to support reproducible research, benchmarking, and downstream curriculum and labor-market analytics.