utter-project/EuroLLM-9B-Instruct
Text Generation
•
Updated
•
64.7k
•
125
A QLORA and ORPO finetune on your ultrafeedback dataset.
It defaults now more to Dutch, even when asking questions in English (sometimes :) )
https://huggingface.co/ReBatch/Llama-3-8B-dutch
I am surprised there is a (small) improvement on dutch_social and hellaswag with only 200k examples for one epoch. All other benchmarks saw a drop, will have to investigate that.
Great, thank you very much!
We were in the process of translating the original ultrachat en ultrafeedback dataset to Dutch ourselves using permissible models for commercial use.
But now we don't have to. Looking forward to using this!