WC models are heavily biased towards old standards/code samples

#8
by FizzyFolk - opened

Wizard coder despite claiming title of best model still miserably fails to use its knowledge of newer standards/versions/approaches when it comes to actual code, heavily biased towards pre 2022 scenarios, libraries, etc. Even GPT-4 fails in that department and even at level of JS spec.

Most likely dataset age issues possibly.

well, 1995 till now🌚

deleted

may not help you, but in the case of python code i always tell models to use version 3.. so i dont get v2 code or a mixture.

And as others have said ( and not just here ), it wont be 'current' the dataset will cut off at a point in time. Its just the way of things..

Sign up or log in to comment