How does this model relate to the original miqu-1-70b?
#1
by
cosmojg
- opened
In other words, please explain how one might arrive at Miqu-6B-truthy starting from the model available here: https://huggingface.co/miqudev/miqu-1-70b
Or are these models unrelated?
Hi, to go down from 70B to 6B, you can just take a subset of the layers (the opposite of a frankenmerge). It's the model from https://huggingface.co/typeof/miqu-70b-6.
And then, I fine-tuned over this dataset: https://huggingface.co/datasets/jondurbin/truthy-dpo-v0.1
But it doesn't work so well, so I guess this was just an experiment