Llama3 Tokenizer

[Bearsaerker]

What is the issue?

I requanted the llama3 Sauerkraut with the newest release of llama cpp which should have fixed the tokenizer, but when I load the model into Ollama, I still get the wrong output while people using llama cpp get the right one. So I'd say that there is still something buggy in ollama. Here is the Output.
"What is 7777 + 3333?
Let me calculate that for you!

77,777 (first number) + 33,333 (second number) = 111,110

So the answer is 111,110!"

OS

Linux

GPU

Nvidia

CPU

AMD

Ollama version

0.1.32

[MoonRide303]

Yeah, same here. Freshly converted GGUFs of both original L3 Instruct and various finetunes that give proper answer to this question in llama.cpp (b2776) or koboldcpp (1.64) fail when imported into ollama (0.1.32).

[coder543]

Duplicate? #4026

[MoonRide303]

@coder543 Yeah, seems to be the same thing.

[Bearsaerker]

Oh yeah duplicate, Im gonna close this then.