Review on Phi-2

#45
by tusharpaul - opened

Un-satisfactory results! The results were not upto the mark.
The results were straightforward & fast. Couldn't solve any complex problem.

More improvements are needed so that it can perform better & can compete with LLaMA & Intel Neural.

Microsoft org

Could you give an example of complex problem you wanted to solve that phi-2 couldn't do?

Microsoft org

@namankhator : I replied there directly, for the example you showed there is no problem once you use the right prompting format (namely "instruct: ...\nOutput:", and adding "Let's think step by step." for any reasoning instruction).

I'm still curious of the errors that @tusharpaul found.

@sebubeck for gsm8k evals python coding is used in prompts beside Let's think step by step ? i am only getting 52% using Let's think step by step + zero shot

Sign up or log in to comment