Quiz

Simple but can be fun. First you select topic, which can be anything (so let your creativity fly or do something common). You get ten questions generated by LLM. It also judges your answers and gives you points for correct ones. At the end it evaluates your performance. You can free chat afterwards if you wish.

As usual, bigger models are good as they understand the format better, but can work also on smaller modern models (though it is more likely the count will get incorrect). In this case you are probably better served with base/instruct model, not RP merge/finetune. Also lower temperature/more deterministic sampler than usual can be good as it is less likely to mess format/instructions.

Make sure example dialogues are turned on! They contain examples with correct message structures so are quite important for the card to work.

Some models I tested and worked quite well: L3.3 70B, L3.1 70B lorablated, Gemma2 27B it, new Mistral 24B (202501). Unfortunately the only small model I have currently downloaded is 12B-ArliAI-RPMax-v1.2 - and it worked too but quite often forgot to increase score with good answer - but not really a showstopper (live with it or edit score manually).

Enjoy!