what are these experiments?
new kind of dataset? new technique to merge models? new base models or something else?
Hi Jawad! Yam Peleg conducted a series of experiments to find the best training protocol, and these models are the results.
You might want to check the original model’s card 🤗: https://huggingface.co/yam-peleg/Experiment26-7B
He did not describe any experiment. Anyway, is the model any good? I have seen many models going to top but when used for extracting information in json format or try to use them for function calling they fail in a way you can only see the failure after completing system and try to run them on large data (on smaller data they may look to be working fine but given large data they just don't work as hoped for). But how about this model?
I've however found that models trained on cosmopedia are much better at following instructions, thus in your collection of models I think mistral pro 8b is the best model.