Contributing

Repo layout

• Code: src/flash_ansr/
• Tests: tests/
• Configs: configs/
• Scripts: scripts/ (training helpers)

Environment

• Use the flash-ansr conda env; install extras from requirements.txt.
• For GPU work, confirm CUDA is available before running tests.
• Large assets (models/data) are pulled via scripts or flash_ansr install.

Quality checks

• Tests: ./scripts/pytest.sh (full suite). For a fast spot check: pytest tests/test_inference.py. (Baseline tests moved to the srbf package in v0.6.)
• Lint: ./scripts/pylint.sh (respects repo config).
• Formatting: follow existing style; stick to ASCII unless a file already uses Unicode.

PR guidelines

• Prefer small, focused changes with tests when logic changes.
• Keep configs stable; when adding new ones, mirror relative paths used elsewhere and document them.
• For decoding/refinement changes, run tests/test_inference.py and related decoding tests.
• New dependencies go to requirements.txt; gate experimental imports.
• Update docs when adding CLIs or config fields.

Style and docs

• Add succinct comments only for non-obvious logic; keep tensors float32 unless a precision module requires otherwise.
• When adding dataloader fields, update FlashANSRDataset.collate to pad/stack consistently and document it.
• Keep README.md lean; detailed usage belongs in docs/ (API, training). Evaluation, baselines, and benchmarking moved to the srbf package in v0.6 (pip install srbf, https://github.com/psaegert/srbf).

Releasing

• Ensure checkpoints include model.yaml, tokenizer.yaml, and state_dict.pt.
• Sanity-check bundles by loading them via flash_ansr install <repo> in a clean environment.