sjtu-sai-agents/datamaster/operate-tools

base_dataloader.py + "\n\n" + code_{node_id}_dataloader.py + code_{node_id}_template.py

{workspace}/
├── input/                          # Input data (train.csv, test.csv, val.csv, ...)
├── submission/
│   └── submission_{node_id}.csv    # Generated submission file
├── code_{node_id}_dataloader.py    # YOUR code (modifiable)
├── code_{node_id}_template.py      # Parent's template (read-only)
└── code_{node_id}.py               # Assembled code (auto-generated on run)

# No import statements needed - auto-assembled with base_dataloader.py
class MyDataLoader(BaseDataLoader):
    def __init__(self, **kwargs):
        super().__init__(**kwargs)
    def setup(self):
        # Load data, feature engineering, augmentation
        # Must set self.train_data and self.test_data
        self.train_data = ...
        self.test_data = ...
    def describe(self) -> str:
        return "Description of data processing approach"

1. read-code      → View base + dataloader + template
2. write-code     → Write MyDataLoader class
3. run-code       → Execute (auto-assembles all components)
4. fix-code       → Fix issues in dataloader
5. validate       → Validate submission file
6. grade          → Get final score

scripts/
├── operate_cli.py           # CLI wrapper
├── for_datanode.py          # Tool functions (from operate_submission/)
├── _submission_utils.py     # Shared utilities (from operate_submission/)
├── base_dataloader.py       # BaseDataLoader ABC
└── grading_server_runner.py # Standalone grading server