Hundreds of dollars later.
Dozens of failed finetunes.
Sisyphus has balanced his rock on the summit.
One must have imagined him happy while pushing. Now, he is ecstatic.

This is a pretty generic finetune of the 24b base model for multiturn instruct RP. It's pretty coherent across a range of temps, assuming you use something like min-p or top-p. It also supports reasoning blocks.

I tested with the following Claude-like system prompts, however they were not trained in and any similar prompts can likely be used:

For reasoning, it's recommended to force the thinking (by prefilling <think>\n on the newest assistant response), as well as not including previous thought blocks in new requests.

v7-Tekken, same as the original instruct model.

Okay. So.

It was supposed to be trained on the above dataset.

However, my dumb ass. Put. The roleplay data. Into the config. And didn't notice

So. Um. I guess this is supposed to be a roleplaying model? It sure doesn't act the part.