Evaluation/Self evaluation
What is this
This preset attempts to make the AI think each response needs to earn points by treating responses like test scores, and self evaluating at the end. Tested heavily with claude(opus/sonnet 4). Tested lightly with gpt(4.5/4.1), deepseek(v3/r1), and gemini(pro).
Important
- The eval at the end of the response will end up using thousands of output tokens. If this is a problem for you, use the stopping strings. If this isnt a problem for you, use the regex.
- Noass is required.
- Gemini pro will filter on certain bots.
- Caching wont work.
- The CYOA preset isn't actually functional as its auto success, I just like having the choices because I am lazy.
- The preset on default tells the AI to roleplay for {{user}}. If you don't want this, disable the "Roleplay for {{user}}" prompt and remove some lines out of the "Evaluation" prompt.
- If you want to change the POV look in the "Main" prompt.
- To change the response length enable the prompts "Longer Response Prefill", "Shorter Response Prefill", "Variable Response Prefill"
- For any deepseek model the temperature should be lowered from 0.85.
Downloads
- Noass: gitgud
- Noass evaluation preset: catbox
- Evaluation regex: catbox (
<(E|e)valuat(.|\n)*) - Preset (Regular): catbox
- Preset (CYOA): catbox
- Stopping strings: catbox
Contact
anon10101000@proton.me (feedback/help)