Ettin Suite: SoTA Paired Encoders and Decoders — outperforms Llama 3.2 and SmolLM2

AI Impact Summary

The Ettin Suite introduces a new open-data training recipe for paired encoder-only and decoder-only models, achieving state-of-the-art performance and surpassing established benchmarks like Llama 3.2 and SmolLM2. This represents a significant advancement in open-data language model development, offering a directly comparable architecture for both encoder and decoder tasks, and highlighting the architectural advantages of encoder models for discriminative tasks.