Information on the model
Hey!
I am interested in this model.
On which dataset it was trained?
Just SFT or also some form of alignment?
Ciao @anakin87 , I follow you everywhere what a honor!!!!! This is a failed experiment because the training was before the gemma fix on HF transformer but also on many other fine tuning lib. The Gemma integration was broken. So it does not work as expected. In any case the dataset used was https://huggingface.co/datasets/mii-community/ultrafeedback-translated-ita, and no DPO or alignment.
I have a plan to redo the experiment in the next weeks.
Thank you!!!
Can you tell me what was the problem with Gemma?
Yes, it is explained very well here https://unsloth.ai/blog/gemma-bugs and addressed in part here https://github.com/huggingface/transformers/pull/29402. I tried the sft before the fix and the result was a mess.