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·May 19·5 min read
Introducing the Ettin Reranker Family
TL;DR
Retrieval systems often return many candidates, but ranking all of them is expensive. Hugging Face released six Ettin Reranker models—specialized neural networks that re-score retrieved results to improve relevance—trained via knowledge distillation from larger models.
✦ Why It Matters
Engineers can deploy production-ready rerankers or quickly fine-tune them on domain-specific data using published recipes and tooling.
Key Takeaways
- Use the Ettin Reranker models for improved document relevance scoring.
- Leverage the retrieve-then-rerank approach for efficient processing.
- Experiment with different model sizes to balance speed and accuracy.
- Install and fine-tune models using the new train-sentence-transformers skill.
Full Summary
The Ettin Reranker family introduces six new models based on ModernBERT encoders, designed for reranking tasks in information retrieval. Rerankers, or pointwise cross-encoders, evaluate the relevance of document-query pairs by allowing them to attend to each other during processing, resulting in more accurate scoring compared to traditional embedding models.
The models were trained using a distillation recipe on a curated dataset, achieving impressive performance metrics on benchmarks like MTEB and NanoBEIR. For instance, the 1B model closely matches the performance of a larger teacher model while being significantly smaller.
These models can be easily integrated into existing retrieval systems, enhancing the quality of search results without a substantial increase in computational cost.