>>73949961People on /lmg/ were having some luck with the 8b model. Are you testing that or the 70B? Apparently you need to make sure that 'ignore special tokens' is still enabled.
I also wanted to confirm that you're using the official template here: we're still trying to figure things out anyway and its still a bit split with old alpaca.
https://llama.meta.com/docs/model-cards-and-prompt-formats/meta-llama-3/https://desuarchive.org/g/thread/100066834/#100068130