Samplers
Why such a heavy rep and presence penalty is required?
These samplers naively apply to ALL of the tokens in the context.
About the reason, if you continue to scroll down, you'll find the same information again but with a bit more explanation, which are quite obvious but nervertheless.
Otherwise, I can answer you about the repeat penalty: 1 actually disables it, at least for transformers, vllm, sglang and llama.cpp:
https://deepwiki.com/search/what-is-the-value-to-disable-r_daace7a8-f367-4fbb-8270-85bc8a53f0ee
https://deepwiki.com/search/what-is-the-value-to-disable-r_be2915a0-e089-48ec-8190-4a85f7780a3f
https://deepwiki.com/search/what-is-the-value-to-disable-r_e86c8fec-2f85-4894-b1e5-285f2c987018
https://github.com/ggml-org/llama.cpp/tree/master/tools/server#sampling-params
Owao is correct. Rep penalty 1.0 turns this particular sampler 'off'.
As for why presence penalty is on at all (let alone 1.5) is a bit of a mystery and it seems to cause problems of its own so I can't imagine it was something they wanted to be necessary.
What I can tell is that I indeed have loop issues when setting presence_penalty to 0 + temp 0.6 on general long reasoning tasks (for which they instead recommend presence_penalty 1.5 + temp 1). Using the latter, I never observed loops, and it works really great. So I'd personally trust them and follow their recommendation to the letter!
By the way, what an awesome model! Surprises me every time
Yes it really is impressive, it (27b) is better than Opus 4.6 for vision tasks!
It makes sense for general reasoning to need temp 1, that seems to be common for the older reasoning models so a model can get itself out of a loop and I guess presence penalty is also doing the same job here. I'm curious though if it gets into loops on coding tasks with it turned off or not.