Ettin

"Seq vs Seq: An Open Suite of Paired Encoders and Decoders": the first collection of encoder-only and decoder-only models trained from scratch with identical data, architecture, and recipe — six sizes from 17M to 1B, each in both flavors, on 2T tokens of open data (DCLM, Dolma v1.7, code, scientific papers). A clean testbed for the encoder-vs-decoder question: encoders win MNLI even against larger decoders, decoders win generation, and cross-objective continued training doesn't close the gap. JHU-CLSP (Weller et al.), MIT-licensed with released data and checkpoints.