Fine tuned with my relatively small (about 3000 samples) sample of roleplay conversations. this datasets RP conversations have 3-15+ turns. this is my $60 attempt to force qwen3 30b to be able to handle RP stuff.
3 epochs, 4e-4 learning rate, because screw you qwen3...
more info after I can test it.
Q4_K_S gguf incoming soon.
Reasoning -
here is a cool ST setting that as processed good results.
image/jpeg
<think> Alright, my thinking should be concise. What are the top 5 things I should keep in mind about the current scene?
1. **
its great because the bulleted lists or already an effective use of tokens, AND you can somewhat control the length of the response by changing the number.
