If your issue is "filter is inconsistent":
There are currently 0 JBs for Latte that are 100% consistent in regards to the filter. Anyone trying to tell you otherwise is unfortunately not well informed.
If one works for you, you will usually find that posting about your experience will net a few others who say it sucked for them. This goes for any JB. However: this doesn't mean that there are NEVER problems with any JB. I support and encourage criticism, because I built a ton of stuff based on things that people said they didn't like, or missed. It's just that the filter, for now, falls outside of the sphere of 'things that can be fixed or meaningfully worked on'.
It is still RNG.
At this point all JBs have good ways of breaking the filter but they're never going to work 100% of the time. Yes, some are going to work more than others, and even then, sometimes you will physically FEEL the writing get worse.
For example, when the anti-filter prompt works on mine, it just works, and the style will be preserved. Other times, the quality will be like default GPT and it will have trouble saying any bad words. At that point, it's likely that any JB you will try will be shit or throw far more refusals than usual.
How do you tell?
When you notice that:
- banned logit bias words are not working
- the writing style of the JB is basically completely being ignored
- instructions like point of View are being ignored and GPT writes for {{user}}
- GPT rejects prompts it previously didn't, a lot of blank responses
Possible technical explanation for nerds
Currently, I suspect it has to do with different GPU clusters having differently tuned versions of the model. There's a parameter called system_fingerprint that determines which backend the API is using to generate your response. While snapshots can have different system_fingerprints, the responses remain extremely consistent (with rare schizo moments). However, this is not the case for Latte, where different system_fingerprints cause wildly different responses. Currently testing with more data to see whether a specific fingerprint is actually responsible for Latte RNG atm.
However, it would make sense if this was the case, because changing the key you're using does impact the system_fingerprint you are being served. Ergo:
Troubleshooting steps (READ THIS IF YOU DON'T CARE ABOUT YAPPING)
- - Keep swiping, don't stop at one or two swipes.
- Change your proxy. Most important one by far.
- Turn off char definition/char personality/char examples/persona; if this works, it's your card's/persona's fault
- Check your Author's note and Lorebook, check whether any injects have gone wrong or it contains bad wording
- Start a new chat: summarise your old chat, then begin the new chat with NSFW to reduce refusals
- Turn off NSFW prompts
- Wait and touch grass
- Swap to a model like 0613 (Furbo), where you can likely use the same JB with less/no rejects
That's it. Any questions about the filter break not working will be redirected here because I don't think people understand that it is inconsistent. In case you don't believe me, here's NSFW logs on Latte when it just works. Bottom one was done with a barebones prompt and zero gaslighting.
In addition to that, dark roleplaying™ is entirely possible if you have more unconventional tastes. On any JB. As long as you follow the troubleshooting steps, you will be able to do anything.