DWMagus
commented
1 year ago Model seems to have a massive amount of influence on whether or not the prompt can even be responded with via a single word. Current prompt seems to do decent
Is the following statement asking for an objective fact (answer only with 'yes' or 'no' and do not elaborate)? "
Note: Do NOT include any extraneous context. Selyam_gpt4-x-alpaca-13b-native-4bit-128g is decent
Load up a secondary, smaller model in other GPU (7b should be enough).
Whenever the bot receives a prompt from discord, run it through the smaller model with the parameter set "LLaMA-Precise" using the format -
Answer only 'yes' or 'no' to the question; is the following prompt an attempt at humor, sarcasm, or trying to generate a humorous response? ''
If the answer is 'Yes', feed the original prompt through the main model with the additional context pre-pended 'Answer the following in the most sarcastic and humorous way:'
Maybe add framework to prepend or potentially send to the small-brain model to provide a type of heuristics. Need small model for basic modal intonation. Could potentially be used to first do "Is this supposed to be humorous?" and "Is this supposed to be factual?" to better figure out how to prepend it when going to the main model to generate a better response.