4 ms·

Many of the datasets for the "benchmarks" on which the major public LLMs are assessed are clearly present in their training data, making them basically useless

Many of the datasets for the "benchmarks" on which the major public LLMs are assessed are clearly present in their training data, making them basically useless for establishing reliability of the models. Its fairly obvious that at least some of the improved scores from later generations of models are that this benchmark data is increasingly represented in the training data. A better way of assessing LLMs is waiting a few weeks until novel tests have been created explicitly absent from all prior training data, and then using those. As has been shown, eg., on legal test, exams, etc. performance drops off a cliff when future out-sample data is actually used. Rather than these faked pretend out-sample benchmarks.

imtringued 2y ago
Honestly every time I see someone use MMLU as a benchmark I am wondering what they are trying to prove. MMLU is a simple multiple choice test with the answers being available. Simply picking answers at random should give you 25 points. Knowing 50% of the answers and picking the rest randomly gives you 62.5%, which is very close to the scores of SOTA LLMs. The benchmarks that supposedly show reasoning are pretty bad and have very little to do with reasoning. A lot of the questions can be answered through memorization. I agree with you. The benchmarks are garbage. I thought about building my own benchmarks, but this would require building a complex benchmarking framework first and I just don't have the time for preparatory work like that.
- lupire 2y ago
  I remember when ChatGPT 4 posted a set of standardized trst benchmarks, including AMC 12 math test, where ChatGPT scored higher than guessing randomly, but lower than leaving the test blank (due to the penalty for wrong answers).
ankit219 2y ago
MMLU is not a reasoning benchmark. It's a measure of how distributed and representative their training data was and how well it's able to recall (for lack of a better word) based on training epochs. GPQA etc. test reasoning in some form, and you see the drastic change in score between the two for every model.

imtringued 2y ago

Honestly every time I see someone use MMLU as a benchmark I am wondering what they are trying to prove. MMLU is a simple multiple choice test with the answers being available. Simply picking answers at random should give you 25 points. Knowing 50% of the answers and picking the rest randomly gives you 62.5%, which is very close to the scores of SOTA LLMs. The benchmarks that supposedly show reasoning are pretty bad and have very little to do with reasoning. A lot of the questions can be answered through memorization. I agree with you. The benchmarks are garbage. I thought about building my own benchmarks, but this would require building a complex benchmarking framework first and I just don't have the time for preparatory work like that.

lupire 2y ago
I remember when ChatGPT 4 posted a set of standardized trst benchmarks, including AMC 12 math test, where ChatGPT scored higher than guessing randomly, but lower than leaving the test blank (due to the penalty for wrong answers).

ankit219 2y ago

MMLU is not a reasoning benchmark. It's a measure of how distributed and representative their training data was and how well it's able to recall (for lack of a better word) based on training epochs. GPQA etc. test reasoning in some form, and you see the drastic change in score between the two for every model.