Here’s the chat fine-tune. I would not have expected such a dramatic difference. It’s just a subtle difference in post-training; Llama 405b’s hermes-ification didn’t have nearly this much of an effect. I welcome any hypotheses people might have.
This looks like what happens when you turn the contrast way up in an image editor and then play with the brightness. Something behind the scenes is weighting the overall probabilities more toward land, and then there is a layer on top that increases the confidence/lowers variance.
This looks like what happens when you turn the contrast way up in an image editor and then play with the brightness. Something behind the scenes is weighting the overall probabilities more toward land, and then there is a layer on top that increases the confidence/lowers variance.