腾讯混元-混元多模态图像理解与标注策略产品（深圳/北京）
Core
Define standards and strategies for captioning and re-captioning multimodal images to support pre-training and post-training of visual language models.
Role type
Senior Product Manager (Multimodal Data Strategy)
Builds
High-quality caption datasets (including subject, attribute, action, spatial relations, and text-in-image) for training large multimodal models.
Domain
Artificial Intelligence / Multimodal Large Models / Computer Vision
Deliverable
production ML models
Required skills
Visual language model strategy, prompt engineering, data quality evaluation, experimental design, cross-functional collaboration
Preferred skills
Experience with image generation, hallucination mitigation, automated evaluation pipelines
Technologies
Visual Language Models (VLM), Captioning/Recaptioning frameworks, Prompt Engineering tools
Responsibilities
Establish captioning standards for image pre-training and post-training; Design prompt strategies and generation workflows for data production; Optimize re-captioning to correct errors and reduce hallucinations; Build evaluation metrics and quality assurance processes; Iterate description strategies based on model performance.