Can't stop. Won't stop.
After responding to user instructions, I notice that this model always volunteers more text, whether in the form of narrative analysis, continuation of the initial text, or postscript, even if instructed to stop. On one run after including the phrase "stop and wait" in the instructions, the model printed "(Waiting)" after finishing the text that satisfied the instructions and then proceeded to ramble on with notes and didactic discourse about the story it generated. In another run, the model printed "The End." followed by continuation text, "The end. (For now.)" followed by a postscript, and "The end. (For real this time.)" followed by a string of emojis and more text inside parenthesis.
In comparison, the Airoboros model included in the merge for this model does not share this behavior. I wonder where this model developed its aversion to stopping.
Regardless, thank you to the creators of this model for building and sharing.
The goal of the model was to be an all-in-one 13b model capable of.. everything, but it did not work as expected. The narration and all is likely from the coder / puddle jumper models being far off compared to the other ones mentioned, and likely something went horribly wrong in the merge. It did well in benchmarks but not real life use cases. Perhaps it would have been better to merge less models at once, then merge the results together instead of doing it in one go. I won't be pursuing this further though, as it's not really my idea in the first place, unless Dampf wants to continue.
It's not a bug, it's a feature. I ask it to explain technical concepts and I love it. It explain in a warm, friendly manner, long-ass explanations with many real life analogies. Without any prompt engineering, at 0.7 temp. I really wish I could get a 70B model to act like this – this one makes factual mistakes which I believe would be more or less improved in a larger one.