Quick Start

Pasteur reads local parquet files and writes a local simulation bundle. This walkthrough exercises all four simulation outputs without uploading data.

Prerequisites

  • Rust 1.95 or newer

  • A parquet file with an ID column and numeric features

  • Optionally, a cohort-membership parquet for flipper simulation

Install the CLI:

cargo install pasteur-cli
pasteur-cli --help

Run a simulation

Choose the feature whose missingness and measurement noise you want to test:

pasteur-cli simulate \
  --input patients.parquet \
  --id-col patient_id \
  --feature glucose \
  --blackout-companion glucose_measured \
  --blackout-rate 0.2 \
  --jitter-iters 3 \
  --output ./output

The companion option keeps the value and its observation flag consistent: when glucose is blacked out, glucose_measured is nulled on the same rows.

Add flipper pairs

Flipper requires a label-membership table. Select the positive cohort by its group_id:

pasteur-cli simulate \
  --input patients.parquet \
  --labels groups.parquet \
  --positive-group-id 1 \
  --id-col patient_id \
  --feature glucose \
  --flipper-pairs 20 \
  --flipper-steps 10 \
  --output ./output

See Data Contracts for the exact input schemas.

Inspect the bundle

The output is organized by simulation type:

output/
├── clean/clean.parquet
├── blackout/blackout.parquet
├── jitter/jitter_0.parquet
├── jitter/jitter_1.parquet
├── jitter/jitter_2.parquet
└── flipper/flipper.parquet

Score models

Score one or more ONNX models against the bundle. Create the results directory first:

mkdir -p results
pasteur-cli compare blackout \
  --sim-root ./output \
  --labels groups.parquet \
  --positive-group-id 1 \
  --model models/model_a.onnx \
  --model models/model_b.onnx \
  --output results/blackout.json

Walkthrough: choosing between models walks through a full comparison, including exporting models to ONNX, and Reading the results explains every number in the output.

Generate a dataset card

Only for bundles made from public or synthetic data that you intend to publish. Never publish bundles made from patient data.

pasteur-cli card \
  --input ./output \
  --source-dataset local/patients \
  --dest-repo org/pasteur-simulations \
  --description "Local clinical-model stress-test bundle"

This writes output/README.md. It does not create or upload a remote repository.

Next steps