Skip to content

CONTRIBUTING.md

Latest commit

 

History

History
48 lines (37 loc) · 2.31 KB

CONTRIBUTING.md

File metadata and controls

48 lines (37 loc) · 2.31 KB
Raw

Contributing to PRIMED-AI

Local Setup

With pip

  1. Clone the repository.
  2. Run pip install -e ".[dev]" or make install.
  3. Ensure Python 3.10+ and PyTorch with CUDA are installed.

With uv (recommended — faster, reproducible)

  1. Clone the repository.
  2. Run uv pip install -e ".[dev]" or make install-uv.
  3. Ensure Python 3.10+ is active in your uv environment.

Running tests

make test        # runs pytest tests/

Linting

make lint        # runs pre-commit on all files (ruff + formatting)

Data Access

  • MIMIC Data Paths: Ensure you have ORCD cluster access and signed DUAs for MIMIC-IV, MIMIC-IV-Echo, and MIMIC-IV-ECG.
  • NEVER commit raw data. (Check .gitignore for ignored directories like data/, logs/, etc.)

Reproducibility

data/, results/, probes/, and logs/ are gitignored, so the cohort, embeddings, checkpoints, and metrics behind the reported numbers are not in this repository. Only code and configs are versioned here. To rerun anything end to end you need all of:

Requirement Where it comes from
PhysioNet credentialing + signed DUAs MIMIC-IV-Echo, MIMIC-IV-ECG, MIMIC-IV
Echo embeddings MITCriticalData/mimic-iv-echo-jepa-embeddings — gated, PhysioNet credentials required. See docs/embeddings.md
ECG embeddings HuBERT-ECG Parquet on the ORCD pool, or re-extract with scripts/extract_ecg_embeddings.py
GPU + storage ORCD cluster (H200 or equivalent for extraction; probe training is minutes on CPU/GPU)

Runs are seeded (--seed, default 42) and every evaluation script writes its input paths, seed, and model dimensions alongside the metrics, so a result JSON records what produced it.

When you report a number, state whether it came from a run you actually executed and on what data. Never present an estimate, a partial run, or a synthetic-demo output as a measured result — src/primed_ai/failure/demo.py generates synthetic data for harness validation only, and anything it produces must be labelled as such.