Skip to content

Public Benchmark Leaderboard

The public benchmark leaderboard is a static, versioned view of OpenMed's archived synthetic evaluation reports. It is rebuilt without network access whenever the documentation site publishes from master and whenever a v* release tag publishes the site.

The renderer ranks every row by leakage ascending and then recall descending. F1 and recall remain visible as secondary metrics, but a higher F1 can never outrank a report with lower leakage. Tabs separate benchmark suites, and model family headings keep comparisons within the same model class legible.

Public artifacts

The stable publication path is https://openmed.life/docs/eval/benchmark-leaderboard/. Each render contains:

  • index.html, the browsable leaderboard with per-suite tabs;
  • leaderboard.json, the same rows in a versioned machine-readable schema;
  • reports/, canonical JSON copies of the archived reports linked from every row.

Every row records the release tag, benchmark run date, and a sha256:<64 lower hex> reproducibility hash. Older reports may obtain model family and reproducibility metadata from the matching models.jsonl entry, but missing ranking evidence fails the render instead of publishing an unverifiable row.

Source and refresh flow

Only committed BenchmarkReport JSON under docs/benchmarks/ with metadata.synthetic set to true is published. Auxiliary JSON that has no BenchmarkReport fields is ignored, while malformed report candidates and reports without the explicit synthetic marker fail closed. Non-synthetic, private, DUA-restricted, or raw clinical data must never be added to this archive.

To reproduce the release artifact locally:

.venv/bin/python -m openmed.eval.leaderboard \
  --reports-dir docs/benchmarks \
  --output-dir docs/eval/benchmark-leaderboard \
  --manifest models.jsonl \
  --release-tag vX.Y.Z

The renderer sorts input paths and JSON keys, derives no wall-clock timestamp, and writes canonical report copies. Re-running it with the same inputs is therefore byte-identical and does not require a network connection.

Machine-readable contract

leaderboard.json uses schema version 1. Its top-level rows array is the canonical global ranking and declares the sort keys as leakage:asc and recall:desc. The suites view repeats those rows under deterministic suite and model-family groups for consumers that mirror the HTML navigation.

Each row includes:

Field Meaning
suite Benchmark suite shown as a tab.
model_family Manifest or report model family.
model_name / device Evaluated model and runtime.
leakage Primary lower-is-better ranking metric in [0, 1].
recall / f1 Secondary higher-is-better metrics in [0, 1].
release_tag Release that published or originally stamped the report.
run_date ISO YYYY-MM-DD date derived from the report timestamp.
reproducibility_hash Validated SHA-256 reproducibility identifier.
report_url Relative download URL for the underlying report JSON.