We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
Hugging Face Blog·29d ago·5 min read
PRX Part 4: Our Data Strategy
TL;DR
To train the PRX model effectively, a diverse dataset was assembled from public and internal sources, focusing on breadth over aesthetic quality. Long, accurate captions were prioritized to enhance the model's understanding of visual concepts.
✦ Why It Matters
Engineers should prioritize using long, descriptive captions in their datasets to enhance model training outcomes.
Key Takeaways
- Focus on dataset diversity over image perfection during pre-training.
- Use long, detailed captions to improve model output quality.
- Leverage Lance for efficient dataset management and exploration.
- Train on JPEG images at quality 92 to minimize storage without quality loss.
How It Works
The data pipeline combines public and internal datasets, using a VLM to generate long, detailed captions that enhance the model's understanding of visual concepts. By prioritizing breadth and diversity, the model learns a wider range of visual attributes, which is crucial for effective training.