Training set for GRPO (Group Relative Policy Optimization) to teach a 7B
model to detect Python lint violations from natural language descriptions.
Filtered large violation spans: Violations exceeding 10 lines or 50 characters
are excluded. This is because future experiments will ask the model to output
the exact violating text in its completion, so violations must be short enough
for a model to reproduce verbatim.… See the full description on the dataset page:
https://huggingface.co/datasets/tumbric/lint_grpo_dataset_v2.
