For leaderboard submissions, see the submission guide. For code changes, describe the problem and how to reproduce it in your issue or pull request.
git clone --single-branch --branch main https://github.com/MachineLearningLifeScience/BEND2.git
cd BEND2
python -m venv .venv
source .venv/bin/activate
python -m pip install -e .Keep model frameworks and weights in optional adapters. The core package
should run on a CPU without cluster tools or private data.
Examples live in src/bend2/_examples/. bend2 init copies them from the
installed package. Use portable paths and load weights on first inference.
Keep automated checks few and focused on evaluation and evidence integrity. Exercise behavior changes with a small example. Build and install the wheel when changing package contents, and check documented commands against the code.
Use short, factual prose. Comments should explain constraints or non-obvious decisions; omit comments that restate the code.
Define each task's scoring in its task module, shared by evaluation and evidence review. Review code independently validates the cohort, labels and submitted evidence before accepting the shared scorer's result.
Increment the affected entries in TASK_VERSIONS in src/bend2/catalog.py when
changing a task's inputs, splits, readout, selection rules or scores. A shared
readout change needs an increment for every affected task. Formatting, comments
and refactors that preserve results keep the same version. Explain whether
results remain comparable in the pull request.
Keep experimental labels out of model adapters. Use predictions to select candidates before reading their measured outcomes. Record checkpoints, training data and source attribution.
Develop on branches based on main. The historical paper-review snapshot
remains in the original research repository.
Keep generated runs, checkpoints, job scripts, manuscripts and experiment archives outside the repository. Link to them when needed for review.