This is copied from the `codeparrot/apps` which is not in Parquet format (meaning that if you are using Datasets>=4.0.0 you will fail to download it because it requires remote code). You can find the origin dataset here: https://huggingface.co/datasets/codeparrot/apps You can find the conversion code here: https://gist.github.com/4gatepylon/024853a9d279812e1f14be93242b3ef8#file-gistfile1-py-L1 NOTE that some of the solutions/input-output are empty. You can check as the code ^ does by looking for `len(element["solution"].strip()) == 0` and/or `len(element["input_output"].strip()) == 0`. Otherwise, these objects are JSON-decodeable strings which look, respectively like: ``` solutions: list[str] # each string is a python function/code you can run that would solve the problem ``` and ``` input_outputs: dict[str, list[list[str] | str | Any]] # keys: "inputs" and "outputs" have same length # usually the list items are lists of strings or are single strings # (I haven't checked exhaustively) ``` Auto-generated docs below: ``` features: - name: problem_id dtype: int64 - name: question dtype: string - name: solutions dtype: string - name: input_output dtype: string - name: difficulty dtype: string - name: url dtype: string - name: starter_code dtype: string splits: - name: test num_bytes: 1226206337 num_examples: 5000 - name: train num_bytes: 103144035 num_examples: 5000 download_size: 788823098 dataset_size: 1329350372 ```