[✅ Solved] Issue with chat's that have specifically formatted responses.

[vivid shuttle]

Thanks for describing the issue in detail! This is a known limitation with many LLMs, especially when they encounter certain formatting cues like headings (***Monika's Thoughts:), list outlines, or brackets that the model interprets as a place to start a new section or stop generation. Here are some steps you can try to improve the behavior on ChubAI with CosmosRP-3.5 (vision):

Suggestions to Fix Response Cut-Off

1. Formatting Change:

   • Try rewording or reformatting the heading in your character description to see if the model responds differently. For example, try:

     Monika thinks, "____"
     

     Or:

     (Monika's inner thoughts: ____ )
     

   • Avoid bold/italics and triple asterisks if the model often halts after those.

2. Prompt Style Adjustment:

   • Instead of having the thoughts start on a new line after the label (***Monika's Thoughts:\n), try keeping it on one line, e.g.:

     ***Monika's Thoughts: She feels nervous and unsure what to say.***
     

   • Models sometimes treat headings ending with a colon and followed by a line break as a stopping cue.

3. Switch the Style

   • Let the model know in the character description how you want thoughts embedded, but demonstrate with a full sentence immediately following the heading:

     Example: Monika's Thoughts: I don't know if I can trust them...
     

   • This helps the model pattern-match and keep its generations going.

4. Temperature Tweaking:

   • If you’re using a lower temperature, try increasing it to recommended (0.8 for CosmosRP-3.5 on ChubAI). Sometimes a very low temp can make the model more rigid with formatting and more prone to stopping early.

5. Try a Different Model:

   • Some models handle formatting quirks better. If you have access, see if CosmosRP-2.5 handles the formatting with less truncation.

6. ChubAI-Specific Note:

   • ChubAI sometimes struggles with longer or complex responses, especially when the API is under load. If the cutoffs are inconsistent, it might be related to latency or slow response parsing.
   • Try using another platform (like VenusChatAI or JanitorAI) as a test to see if the formatting problem is model-related or frontend-related.

Final Notes

• This behavior is not a bug but a consequence of how LLMs see and parse formatting.
• Simplifying or tweaking the structure as suggested often yields better results.
• If testing all the above doesn't work, you may want to try a new chat to clear out old instructions, as custom prompts can persist in the chat history.

If you keep seeing issues specifically with this formatting, let us know in the feedback channel—your feedback helps guide improvements!

Please note: This is an automated response from @vivid shuttle. While it strive to provide accurate assistance, It may occasionally make mistakes. If you find any inaccuracies or need further clarification, feel free to wait for our community helpers to give you further guidance. Thank you.

[✅ Solved] Issue with chat's that have specifically formatted responses.