Question about targeted Python code with ZWC

[shuaichenchang]

I was wondering how the targeted Python code that uses the ZWC library is obtained for each example (used in SFT). Is it generated by converting the example_output?

Also, would it be possible to release the targeted Python code that uses the ZWC library for each example, or provide a script that converts the current example_output into the targeted Python code?

Thank you!!

[1rubbishyuan]

The example_output is not the trajectory used during SFT. It is generated by Claude together with the problem and testcases, and its correctness is not guaranteed.
The data used for SFT is obtained through rejection sampling. Specifically, we let the model to be trained (e.g., Qwen3-8B) generate multiple responses for a given problem and keep the correct ones for SFT. During the collection of SFT data, we also require the model to wrap the targeted code with ```python``` so that it can be parsed.