New model?
I've been using PersonalityEngine for a while now, and I must say that no matter what other models come out, either vanilla, abliterated, or even trained for RP, I have not yet found a model that comes even close to v1.3 for me. It is a truly wonderful model: creative, follows character personality well (as described in the label!), and given it is a 24B model it is pretty fast on my machine.
If no other models ever came out, this would be my top 1 model for RP or even for massive data generation of RP like datasets. That being said, I'm curious whether you're planning a new version in the PersonalityEngine family, or even a new model with different capabilities? I have no expectations of such things, just curious!
Thank you again for such a great model!
For now all personal releases are on hold for the foreseeable future. I'd keep an eye on the stuff pushed out by Arcee AI in the coming months, they might just be what you're looking for.
Oh man, i really do hope that you haven't stopped developement. As aimeri just said, nothing else comes close to your work, with that said, i'll admit that i'm a newcomer to the AI world and i haven't tried that many models yet, but all those models i tried (quite some few) misbehaved, were just plain dumb, destabilized at low ctw windows, etc, broke output formatting and whatnot. Then i stumbled over your model, your work, and "HOLY SH..!" The holy grail of what i have been looking for! Thank you, thank you, thank you!!! :) I Do hope you will continue to dev. I am going to dev a bot using your model, for private use, an evolving bot, and i hope your work is not "dead" in the time coming. Cheers man! You rock!
LeroyDyer/Mult_AgentVerse :
i just made this for role play and chat room based enjoyment !@
its only a app.py and you would need to download it as its not connected to a model online only your own lmstudio !!
but i made 100 agents !! lol !! and a multiagent chatflow !! .. you can pick your choice or create on the flyy characters :
personally i enjoy training the model for roles now !!
I've been using PersonalityEngine for a while now, and I must say that no matter what other models come out, either vanilla, abliterated, or even trained for RP, I have not yet found a model that comes even close to v1.3 for me. It is a truly wonderful model: creative, follows character personality well (as described in the label!), and given it is a 24B model it is pretty fast on my machine.
If no other models ever came out, this would be my top 1 model for RP or even for massive data generation of RP like datasets. That being said, I'm curious whether you're planning a new version in the PersonalityEngine family, or even a new model with different capabilities? I have no expectations of such things, just curious!
Thank you again for such a great model!
im struggling to get anything satisfactory out of this model, are you using the recommended parameters or have you tuned your own?
im struggling to get anything satisfactory out of this model, are you using the recommended parameters or have you tuned your own?
I think as long as you use the SillyTavern Template provided with all the recommended settings, you should be good to go.
I used these settings and the response_format field in llama_cpp_python: https://llama-cpp-python.readthedocs.io/en/latest/#json-schema-mode with my desired response schema.
Didn't really have to do anything special. It must be noted that my prompts were pretty specific to a high degree, but simple in nature. Something along the lines of:You are an expert writer specialized in intricate storytelling. Examine the current chat history and using the provided schema return a complete assessment of the story, the story beats, the scenario, and characters so far.\n\n# Current Chat Log:\n{chat_log_string}\n\n# Schema:\n{ScenarioProgress.model_schema()}
If temperature 0.8 generates garbage output, try reducing it to 0.6 or 0.4. I've found that the lower the temperature, the more consistent I could get responses. It's worth noting that even with temp 0.4 I don't get 100% success, more like 70% of the times it generates a response correctly. Since my usecase was synthetic data generation, and I have more time than sense, I slapped the generation in a loop to check that the response is a valid model (pydantic model, not llm model), and that catches the few times when the response gets cutoff short or something like that.
This model is perfect for my usecase in that it is very creative and doesn't have a positivity bias that is too strong, so with the right setup around it, it is perfect for slowly generating datasets. I also used this model in sillytavern to play with some cards, but I am yet to find a model that is small enough to fit in 32gb of unified memory (no ram+vram tricks for me) and smart enough to follow a long-horizon multi-turn conversation. Bigger models seem to be able to do that, but not what I am coming to this one for.