Thank you
This model is perfect, amazing, like a chef's speciality dessert. The description says everything: the behavior of the base model was kept with its intelligence and awareness, but it's better for roleplay.
Just out of curiosity, would you try to do the same fine-tune on glm 4.5 (not the air version)? I'd gladly do it myself but I never fine tuned a model, so if you have good links/tutorials...
Sorry if my english is bad, I'm not a native speaker.
Hey, glad it's working well for you!
Unfortunately big the version is a bit out of my league in terms of hardware. MoE's are very finnicky and difficult to train due to the tools available currently, so I wouldn't be confident tuning anything bigger than air (also I wouldn't be able to run it myself).
I personally use axolotl to do my finetuning. If getting started I'd recommend starting with a small, well known model like mistral small, Llama 3 so you can trial and error your way through setting up an environment, learning what the config options do etc. It's my personal opinion that the actual process of finetuning for most models (MoE's like GLM-4.5 excluded) is actually pretty straight forward if you're willing to spend the time getting over the initial learning hurdle, but the hardware costs and datasets are the real limitations.