This baseline is intentionally organized for high editability, not maximum framework cleverness.
Objective logic is isolated from trainer wiring in src/hackable/objectives.py.
Rewards are plugin functions in src/hackable/reward_plugins.py.
Data loading is a provider plugin in src/hackable/data_plugins.py.
The trainer entrypoint (src/train_grpo.py) only composes modules from config.
You can load an external… See the full description on the dataset page:
https://huggingface.co/datasets/psidharth567/neuralese-move.
