Pretty good in narrative, but...
After testing the model for a while, I really like the way the model narrate. However, it also seems to have lost too much intelligence in my opinion.
Like the model page says, it understands quite well the character's personality, but it is having trouble figuring what to do next. The model also doesn't seem to like taking the lead, and will often leave its responses open ended.
It also has a tendency to write for the user. Although I've mitigated this in SillyTavern by disabling the Character's name in sequences (Something that, incidentally, I have found greatly improves the responses of Drummer's Behemoth 123B based on Mistral 2411 as well, which is currently the best finetune of Mistral 2411 I've found so far, despite not having liked Behemoth 123B based on Mistral 2407 before. I wouldn't have thought about trying that if this model didn't point me towards disabling that).
In the end, I think this model is strong in narrative, but unfortunately stands a bit short on everything else. It's a model I'd like to try and merge at one point to try and get the narrative, and character sheet understanding, but on its own, unfortunately, performs poorly in a scenario.
Wish you luck on your next finetunes !
ty for the feedback <3
After testing the model for a while, I really like the way the model narrate. However, it also seems to have lost too much intelligence in my opinion.
Like the model page says, it understands quite well the character's personality, but it is having trouble figuring what to do next. The model also doesn't seem to like taking the lead, and will often leave its responses open ended.It also has a tendency to write for the user. Although I've mitigated this in SillyTavern by disabling the Character's name in sequences (Something that, incidentally, I have found greatly improves the responses of Drummer's Behemoth 123B based on Mistral 2411 as well, which is currently the best finetune of Mistral 2411 I've found so far, despite not having liked Behemoth 123B based on Mistral 2407 before. I wouldn't have thought about trying that if this model didn't point me towards disabling that).
In the end, I think this model is strong in narrative, but unfortunately stands a bit short on everything else. It's a model I'd like to try and merge at one point to try and get the narrative, and character sheet understanding, but on its own, unfortunately, performs poorly in a scenario.
Wish you luck on your next finetunes !
Can you elaborate on where you disabled the character's name? Was it in the DRY sequence breakers? In the custom stopping strings? Thanks!
Can you elaborate on where you disabled the character's name? Was it in the DRY sequence breakers? In the custom stopping strings? Thanks!
It's in the AI Formatting tab, in Context Formatting. I have found that unchecking the first checkbox (And as such removing the character's names before a response), improves the response quality. I am using the pure Mistral Template with System Prompt Token (Basically, the instruct template of Mistral 2411, no Metharme hybrid or anything like that.)