Train NousResearch/Hermes-2-Pro-Mistral-7B on the Internal Knowledge Map dataset

#4
by Joseph717171 - opened

Idea: Use the Internal Knowledge Map dataset to Train NousResearch/Hermes-2-Pro-Mistral-7B
Observe how your dataset is assimilated into the model, and see how it changes how it thinks. I'm curious how your dataset and the behaviors it causes to emerge change other llms train on even more diverse datasets. 🀩

https://huggingface.co/NousResearch/Hermes-2-Pro-Mistral-7B

Sounds like a great idea! Since this is such a popular model, this might actually be the one to really form a baseline and comparison tests against to see how the dataset affects the outputs. I'll start this right away. Thanks for the input and awesome idea. Also, I'm adding new aspects to the dataset, is there anything in particular that'd you'd like a general chat model like Hermes-2 to have? For example, better Role Playing, better problem solving, more personality? I've found a few ways to expand the dataset into new areas while maintaining the core structure, but I want to make sure that what I am adding is actually useful to a user and not just bloating the LLM

Owner
β€’
edited Mar 14

Edit: something happened with the Unsloth training and safe-tensors so it might not work. It did about 50% of the time for me. All that being said, already working on the improved version! thanks for the great idea!

First version done! https://huggingface.co/Severian/Nexus-IKM-Hermes-2-Pro-Mistral-7B

I haven't been able to fully test it yet or convert it to GGUF (not at my normal workstation today). I'll do more thorough testing, converting and retraining later this evening.

Sorry for the late reply. I am interested in llms ability to reason (like Microsoft's Orca-2-13B) and their ability to RP (roleplay). I look forward to testing your latest creation. πŸ˜‹

Sign up or log in to comment