Multimodal Data Governance
Governance and safety controls for image-text, audio and video corpora: alignment, provenance, PII, consent, safety filters and deletion workflows.
Что должен уметь кандидат
- Explain how multimodal datasets add alignment, safety, storage and rights-management issues beyond text-only corpora.
- Treat CLIP-style filtering/alignment as curation signal, not proof of safety or legality.
- Specify dataset-card fields: sources, collection process, filters, intended use, biases and unsafe-content risks.
- Propose governance workflows for exclusions, audits and downstream user warnings.
Что спрашивают на собеседовании
- How would you curate a web-scale image-text dataset responsibly?
- What metadata does video need that text does not?
- How would you handle remove requests after model training?
Практическая задача
Create governance checklist for image-text/video dataset: sample schema, safety filters, audit sampling, dataset card and escalation process.
Source-grounded правило
Safety scores and filters are imperfect signals; publish limitations and review process rather than claiming guaranteed safety.