Would love to collaborate and use this dataset even more
so i made a cool mistral fine tuned model, i was hoping we could collaborate to do more, maybe some demos that are interesting but most of all and especially on the evaluations and benchmarking. in case you missed it , here's a deployment of fine tune based on your dataset : https://huggingface.co/spaces/pseudolab/MistralMED_Chat
That's great! Actually, I tried fine-tuning on Mistral model, but that did not impress me enough. I'd definitely love to check your space.
Also, let's think about how we could improve it further to be more conversational + past context driven + improved the user experience of using this utility.
PS: I'm yet to check your space. I'll review it in the next comment here. Many thanks for sharing it.
Hey
@Tonic
I recently checked your HF Space MistralMed_Chat, the development is impressive, however, on a positive note I can see the scope for improvement in terms of grammatical errors and response relevance. Let me know if you wish to work further in this, we could build something good with the dataset.
yes ! let's for general purposes , there are plenty of great conversational, multilingual, generalist datasets , but the best would still be to benchmark the performance first.
I'm also a big fan of epochs over n_steps , so several fine tunes in sequence also seems interesting .
also maybe more steps should work, since there is so much data, so there should be so many steps also - maybe a re-training is in order, indeed .
Yes, actually I am still at a limited (free version) of Google Colab T4 GPU which has several time and disk space limitations. I am definitely considering retraining and improvements in order to make the model work efficiently.
Also, I am quite new to HF Spaces so unsure how to work around with a lot of stuff. I am still learning. The space seems fun!
so a nice little open source friendship was born ^^
