PRX Part 4: Our Data Strategy
PRX's data strategy focuses on building a diverse, high-quality training dataset by combining public and internal sources, using re-captioning with a VLM to ensure accurate descriptions, and storing the data in Lance for efficient querying. The pipeline includes steps for filtering, deduplication, and compaction, with JPEG images encoded at quality 92 being used due to their practicality for training. The dataset is processed with Ray Data for performance and efficiency, and the use of a VLM for captioning significantly improves model performance. The dataset is structured to support both training and inference, with the final model trained on a large, diverse dataset that captures a wide range of visual content.
More from Huggingface
LFM2.5-VL-3B for Better and Faster Vision Capabilities for the Edge
Thinking of ACE? We Can Do It with Fewer Tokens
Meta is back with Muse Glimmer: local, agentic, multimodal, and open source
