>>108527470
>>Is there a standard prompt to test it with? What's with the pixel maps? Is that some sort of function calling?
I just make 450 requests each for one token with this prompt:
I want to know what continent is at the location with given coordinates (or, if there is ocean/sea there)
The coordinates are: latitude={lat}° and longitude={lon}°
Answer with 1 if land and 2 if ocean.
(the last line is approximation since it's generated by code and I don't want to bother looking it up exactly)
And then I look at probability of 1 and 2 in the model's answer using the logprobs argument for c hat completions api.