It’s a very bad use case for an LLM because they don’t have as much polars code in their training data and the polars API has changed a lot in the past couple of years.
Copilot is still suggesting out of date polars code to me. ChatGPT and Claude regularly suggest operations that are valid in pandas and not in polars.
Can confirm this, LLMs write almost exclusively in broken polars. However I convert pandas files to polars semi-regularly, and it's still a hell of a lot faster to get an LLM to write the first draft and correct the 4 or 5 attribute errors it makes
Copilot is still suggesting out of date polars code to me. ChatGPT and Claude regularly suggest operations that are valid in pandas and not in polars.