Differences between 1.0 and 1.1?
Any details about the finetuning? What was changed that makes 1.1 better than Tess-M 1.0?
I would also like to know!
Mentioned by https://huggingface.co/migtissera/Tess-M-Creative-v1.0 that V1.1 deprecated v1.0 models, possibly achieving better in both creative and STEM the same time?
@brucethemoose
do you have plan to merge this model with Capybara?
Capybara-Tess is my favorite model so far!
@brucethemoose do you have plan to merge this model with Capybara?
Capybara-Tess is my favorite model so far!
That's good! I havent really gotten much feedback about the merge.
Yeah, probably soon once I can test Tess M 1.1? I am hoping another good 200K finetune comes out to add in.
Hey guys,
Tess-v1.1 has a lot more data. I'm still creating the complete dataset for Tess, but wanted to see how things work out at different levels. Tess-v1.0 was trained according to LIMA paper (~4000 samples), but the mistake was that it was only trained for 1-epoch. Hence why the issues with instruction following. Tess-v1.1 has a lot more data and more training epochs. I'm still doing R&D so things will keep improving as I release newer versions.
Sorry guys, v1.2 will be coming out soon as well -- Probably tomorrow. You have to put up with frequent releases for the next little while. :)