Loading post…
Hugging FaceTom AarsenAnnouncement
Introducing the Ettin Reranker Family
Summary
Tom Aarsen introduces six Sentence Transformers CrossEncoder rerankers, ranging from 17M to 1B parameters, alongside their training data and full recipe. Each model uses an Ettin ModernBERT encoder with unpadded attention, RoPE, GeGLU, up to 8,192 tokens of context, and a modular classification head whose ablations favored CLS pooling over mean pooling. Training applies pointwise MSE distillation to scores from mixedbread-ai/mxbai-rerank-large-v2 using a broad and retrieval-focused dataset mixture. The conclusion reports state-of-the-art results at every released size, with the 400M model within 0.0024 and the 1B model within 0.0001 of the 1.54B teacher on MTEB, while all models outperform the ms-marco-MiniLM-L*-v2 family on MTEB and NanoBEIR.
Context
Cross-encoders provide more accurate query-document relevance scoring because the two texts attend to each other through every transformer layer, but running one for every corpus pair is expensive. The stated production pattern is retrieve-then-rerank: an embedding model selects top-K candidates, after which a cross-encoder reorders that smaller set.
Approach / What changed
The release builds six CrossEncoder models on six Ettin ModernBERT encoders, using a four-module classification head with Transformer, CLS pooling, Dense, LayerNorm, and final scoring layers. The models were trained through pointwise MSE distillation from mixedbread-ai/mxbai-rerank-large-v2 over cross-encoder/ettin-reranker-v1-data, a mixture of broad-domain and reranked retrieval data, with one shared training script across sizes.
Takeaways
The six model checkpoints span 17M, 32M, 68M, 150M, 400M, and 1B parameters, allowing quality and speed trade-offs within the same reranker family.
All released rerankers accept up to 8K tokens of context, and the post reports 1.7x–8.3x speedups over default loading when using bfloat16 and Flash Attention 2, depending on model size and sequence length.
The released dataset contains approximately 143M query-document-label triples across 39 named splits, making the provenance of individual rows visible.
Source and freshness
- Source
- huggingface.co
- Published
- Added to Yomu
- Last updated
Summaries are generated from the original by a language model. Check the original before relying on it.
Keep reading
Related reading
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
The best engineering reads, once a week
One useful email. Unsubscribe anytime.