Downtown-Case

/

internlm_internlm2_5-20b-llamafied-hacked-rope

like 2

It's the base model.

It should work, but I need to upload a small config change to get it to work in certain UIs.

Sorry to ask you another question. What prompt format and sampler settings are you using?

Depends on the app, I either use raw text (like a long novel) or a short chatml prefix (since there are chatml tokens in the tokenizer). If I'm finetuning it, I just use chatml.

It tends to like lower temperature that other models, 0.08 MinP, and a very low rep penality, with DRY if applicable.

To be honest, I've switched to the new Command-R for most usage, as its so much smarter under like 100K, fits on 24G cards very well, and this model gets a little dumb past 128K.