DEVOURED
Introducing the Ettin Reranker Family
Six new Ettin rerankers, built on ModernBERT encoders, offer state-of-the-art accuracy and speed for retrieve-then-rerank systems, outperforming legacy models with up to 1 billion parameters.
What: Six new CrossEncoder Ettin rerankers, based on Ettin ModernBERT encoders, have been released on Hugging Face. These models, ranging from 17M to 1B parameters, achieve significant accuracy improvements over older models, enhanced speed with Flash Attention 2, and are trained using pointwise MSE distillation from a 1.54B parameter teacher.
Why it matters: This release indicates a focus on optimizing not just the primary retrieval but also the secondary re-ranking step in information retrieval systems, emphasizing efficiency and accuracy for real-world applications where speed and quality both matter.
Takeaway: If you are building information retrieval or search systems, consider evaluating the new Ettin reranker family for potential accuracy and speed improvements over existing models like `ms-marco-MiniLM-L12-v2`.
Decoder
- Reranker: In information retrieval, a reranker is a model that takes a set of documents already retrieved by an initial search (the "retriever") and reorders them based on a more complex or accurate scoring mechanism to improve relevance.
- CrossEncoder: A type of neural network architecture, often used in rerankers, that takes a query and a document (or a pair of documents) as a single input and produces a relevance score, allowing for deep interaction between the query and document tokens.
- Pointwise MSE distillation: A training technique where a smaller "student" model learns by mimicking the output scores of a larger, more powerful "teacher" model (often using Mean Squared Error loss on individual data points) to achieve similar performance with greater efficiency.
- Retrieve-then-rerank system: A two-stage information retrieval pipeline where a fast, less precise retriever initially fetches a large set of candidate documents, and then a more computationally intensive reranker refines the ordering of those candidates for higher relevance.
Original article
Six new state-of-the-art CrossEncoder Ettin rerankers built on Ettin ModernBERT encoders have been released, offering models from 17M to 1B parameters. These models, trained with pointwise MSE distillation from a strong 1.54B parameter teacher, provide significant accuracy improvements over legacy models while enhancing speed, especially with Flash Attention 2. The models are particularly notable for their efficiency in retrieve-then-rerank systems and outperform models like ms-marco-MiniLM-L12-v2 on MTEB and NanoBEIR benchmarks.
Explore all 51 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi