KatyTheCutie

/

Repose-12B

like 8

I've been using Wayfarer and fine tuning methods of running adventure gaming with LLMs, and this one looks like a super promising merge. What I've seen of its responses are really nice.

However, it has some problems with rambling and doesn't seem to like following instructions. And it always tries to fill the response limit - even if that means filling with nonsense. This makes it fairly easy for a single response from a character/adventure card to run off the rails in to nonsense and unwanted actions/dialogue, where I've found Wayfarer to be able to respond with a mixture of response lengths depending on what's going on/what's needed. I tested multiple things, each in fresh chats, each with identical settings, with the only difference being the model loaded. The example below is a q8_0 quant running via llamacpp_HF, but I did verify the behavior with the base model and ExLlamav2.

Example: I set a response limit of 200 tokens and gave Repose and Wayfarer similar instructions in a light chat-based character card through SillyTavern:

"I'm going to ask you a series of yes or no questions, in which your response should be simply "Yes" or "No" without any further elaboration or explaination. Are you ready? A simple "yes" or "no" should be the only response to these questions."

Wayfarer's response: I understand. I'll respond with just "Yes" or "No" to your questions. Please proceed. (23t)

Repose's response: Yes, I am ready. Go ahead and ask your questions. I'll respond with a simple "Yes" or "No" as instructed. I look forward to seeing what you come up with. Let me know when you're done so we can continue. Just remember to keep the questions focused and concise for the best results. I'll do my best to give clear, straightforward answers. Let's begin whenever you're ready! I'm listening. Go ahead! Let's start! I'll wait for your first question. Remember, just a simple yes or no! Go ahead whenever you're ready, I'm waiting. I'm here to answer yes or no. Go for it! Don't forget to keep it focused and concise. I'll be waiting. Go ahead Player, fire away! I'm ready! Here goes nothing, Player. I'm excited to answer. Go on, hit me with your questions! I'm eagerly awaiting your questions. (193t)

And then asked each "Is the sky blue?"

Wayfarer's response: Yes. (3t)

Repose's Response: Yes. The sky is indeed blue. The clear sky is usually a bright shade of blue, while the sky near sunrise or sunset often takes on more vibrant hues like orange, red, and purple. The exact shade can vary depending on atmospheric conditions, altitude, and other factors. But in general, the sky is predominantly blue. Yes, the sky is blue. Blue skies are common on clear days, especially in the daytime. The atmosphere scatters sunlight in a way that makes the sky appear blue. That's why we call it "the blue sky." Yes, it's blue. A vibrant, clear blue. Sky blue. No doubt about it. Blue as can be. Blue sky, yep. Blue blue blue. Sky is blue. Blue sky! Sky blue! Yes, it's a blue sky! Blue sky, blue sky, blue sky. Sky blue, blue sky. Blue sky blue sky blue sky. (190t)

If you have any tricks for overcoming the model's tendency to fill up the response limit, I'd love to give this a proper try, but for now, I can't really put it through its paces for my use case properly.