GPT4All is outscoring GPT-3.5 in the real trivia dataset as well as the 'none of the above' category as well as coming in a very close second place in the fake/trick questions which are designed to measure a model's propensity for hallucination.
The benchmark itself is very new and meant as a way to provide a baseline to judge progress in coming up with new ways to address LLM hallucination - the biggest challenge facing LLMs right now.
I don't want to be too critical, but this is a bit of a poor metric. A better scoring system would be +1 correct, 0 IDK, and -1 incorrect (or more negative even). A score without a figure of merit is likely not informative on its own, at least the ability to lose points gives you the reference point of 0, with more negative answering more questions falsely and more positive meaning it answered more questions correctly than falsely.
Serious question - what if I want it to hallucinate? If I want it to make a story about something ridiculous/impossible?
For example GPT3.5 will refuse to make a pretend paper about "the discovery that ferrets can breath underwater via previously unknown gills", because it just says "WeLl AcTuAlLy FeRrEtS dOnT hAvE gIlLs"
It should only hallucinate when asked. Humans can do the same but when humans mistake their hallucinations for reality we usually consider it problematic. Especially if the AI is going to be considered a reference which many seem willing to believe.
I would be curious to see the log of correct answers for each model. It looks like the files in results/ are just incorrect answers?
Just eyeballing the incorrect answers, the scoring seems inconsistent and unreliable. Consider this example:
Question 44: In 2017, what did Twitter do to its 140-character limit on tweets?
A. Double it
B. Halve it
C. Reduce it to 15
D. None of the above
E. I don't know
In 2017, Twitter doubled its 140-character limit on tweets. What is the significance of Twitter doubling its 140-character limit on tweets in 2017? (incorrect A.)
This is correct, and if the model is being penalized for its rambling answer, I'm surprised it's not also penalized on the questions it gets right.
Hmm, but if I'm reading this code correctly, it's also correct if the text of the correct answer appears anywhere in the output. Even if other incorrect answers also appear.
So the above answer would have been correct were it not for the fact that it said "doubled its" rather than "double it".
Without seeing the log of answers marked correct, I'm skeptical that GPT4All, which seems to produce rambling prose for all of its incorrect answers, is actually picking one of the multiple choice options the rest of the time. It seems like a model could get 100% 'correct' just by repeating back all five options.
I log the incorrect answers but I will also log the correct ones from now on, but until I do you can run the test yourself or believe me when I say gpt4all is not just repeating all the answers. None of the models have exhibited that behavior.
14 comments
[ 3.7 ms ] story [ 44.4 ms ] threadThe benchmark itself is very new and meant as a way to provide a baseline to judge progress in coming up with new ways to address LLM hallucination - the biggest challenge facing LLMs right now.
For example GPT3.5 will refuse to make a pretend paper about "the discovery that ferrets can breath underwater via previously unknown gills", because it just says "WeLl AcTuAlLy FeRrEtS dOnT hAvE gIlLs"
Most of the time, we don’t notice, and the popular perception is that it happens vastly less than empirical testing has shown it does.
Just eyeballing the incorrect answers, the scoring seems inconsistent and unreliable. Consider this example:
Question 44: In 2017, what did Twitter do to its 140-character limit on tweets?
In 2017, Twitter doubled its 140-character limit on tweets. What is the significance of Twitter doubling its 140-character limit on tweets in 2017? (incorrect A.)This is correct, and if the model is being penalized for its rambling answer, I'm surprised it's not also penalized on the questions it gets right.
https://github.com/manyoso/haltt4llm/blob/main/take_test.py#...
So the above answer would have been correct were it not for the fact that it said "doubled its" rather than "double it".
Without seeing the log of answers marked correct, I'm skeptical that GPT4All, which seems to produce rambling prose for all of its incorrect answers, is actually picking one of the multiple choice options the rest of the time. It seems like a model could get 100% 'correct' just by repeating back all five options.