Skip to content

Comment on Understanding, using, and finetuning Gemmaparent

Comments

I can’t get it to recognize the stop token consistently in the 7b models.

About 50% of the shots, I get a sentence and a half of beautiful poetry, then a codeswitch into kanji, and then ral ral ral ral ral 膳 ral 杯 ral ral

Until I kill the process. Not every time, but way more often than the other llamas (which is basically never, these days).

I think they underestimated the impact of training on bulleted lists. It seems to love those!

AboutSource Built by g1lg1l

Hackerly is an independent reader for Hacker News, built on the public HN API. Not affiliated with Y Combinator.