Q8_0

#1
by snombler - opened

Hi there.

I test models a lot. Would love to see a Q8_0 full of your hot glue. Thanks.

Thanks for your interest!

The 8_0 (and the fp16 for potential future quants) are uploaded, feel free to take it for a spin.

Thanks! Initial first little few test gens looked fine. I'll put it through its paces here before too long.

snombler changed discussion status to closed

I did some limited testing (sorry, I'm busy lately). So just to throw some feedback in, I feel like the variety of gen length changed a bit and it felt like there was a bit of compression of creativity. Not a ton, but noticeable on the more esoteric cards with odd speech patterns. I didn't notice a large drop in attention to detail but I didn't do a ton of testing, as I said.

I do think it tamed the quirks a bit, but I believe it may have over-tamed and hurt some of the gen variety and word choice. I'd also say take this feedback with a grain of salt because it's based on limited number of chat sessions (~5) with no real strict A/B testing against the original 20B.

Sign up or log in to comment