Tom Aarsen

Thanks Maxime! You know, I bet the training script would apply pretty cleanly on LFM models as well 👀
Although maybe the model prefers a generative architecture with a chat template that asks the model to generate either "yes" or "no", and the difference in raw logit scores for those tokens are used as the prediction. That's what's used for the teacher model's Sentence Transformers integration, and it works quite well: https://huggingface.co/mixedbread-ai/mxbai-rerank-large-v2

Yess, exactly. In my opinion, rerankers have been getting too big to keep them usable. Hopefully this'll inspire others to make smaller models as well. Plus, the Ettin base models are just amazing, that helps a lot.

Definitely! I tried a few variants, and in my tests, a mix of top and stratified worked best, but I sampled from 2048 docs, so fully stratified means the 2nd doc is already pretty far from the most similar.