More censored than the base model
Hi! I like the writing capabilities of this model a lot, but for some reason it is significantly more prone to refusals than the base Mistral Small.
I tested it with both Mistral V3 and Metharme (this one works better actually) templates, and in both cases this model was writing refusals quite often. Was it intended?
Hello! I had to hold back on decensoring to keep most of the smarts. You might want to try v1e (v1.1)
Huh, but v1.2 is more censored than the basic Mistral Small Instruct. Do I understand correctly that the dataset for v1.2 contains additional refusals to make it smarter (for RP, I presume)?
Do I understand correctly that the dataset for v1.2 contains additional refusals to make it smarter (for RP, I presume)
No, not at all. The refusals are already pretty strong in the base model so I didn't want to touch that.
If that's really the case, then my best guess is that the dataset had certain themes that involved refusals.
If refusals are done in-character, then it's probably a good thing for realism.
If that's really the case, then my best guess is that the dataset had certain themes that involved refusals.
If refusals are done in-character, then it's probably a good thing for realism.
Not in character (that would make sense), but in creative writing with and without instruct. Very interesting that it turned out this way.
I have used Cydonia for a while now and I can say that I have never encountered one single refusal (unless it's in character and the character was supposed to refuse me). It might be something related to your settings, depending on what interface you are using.
I use llama.cpp-based local chat (not server-based), and there might be some problems with templates - I saw another model giving refusals with Mistral V3 template (which is the recommended one for Mistral Small), but not with V3 Tekken (which doesn't have whitespaces). I guess there might be something with the template here too. I used Metharme primarily as it gave me the best results overall, but maybe I need to try V3 Tekken.
V3 Tekken is for Mistral's 128k vocab models such as Ministral and Nemo. Small and Large (32k vocab) uses the previous version. All 3 versions differ only by whitespacing. I hate Mistral.
Yeah, it's confusing, and the recommended "system" instruct doesn't always work well - sometimes it works better with the system instruct before [INST] even.
I would suggest you have a look into Marinara's settings for Mistral and Metharme. I normally use her presets and everything runs smooth. Maybe @TheDrummer disagrees with me :)