OPEN DATA · CC BY 4.0

Inspect the data, not just the verdict.

Stable identifiers, field-level provenance and versioned exports make every answer auditable and reusable.

JSON · FULL DATASETDownload v2026.08.18.3Models, hardware, quantizations, sources and runs →CSV · FLAT EXPORTDownload result rowsOne record per observed attempt →JSON SCHEMA · v1Run receipt schemaValidate a candidate observed result →

DATA CONTRACT

Stable, explicit fields

{
  "datasetVersion": "2026.08.18.3",
  "generatedAt": "2026-08-19T03:21:00Z",
  "models": [...],
  "hardware": [...],
  "quantizations": [...],
  "runs": [/* 10 reviewed attempts */]
}

Successes and failures

The production dataset contains 10 reviewed Qwen3.8-27B × RTX 4090 attempts. OOM rows remain first-class evidence, with zero values rather than invented throughput.

EXAMPLE QUERY

Filter without an API key

fetch('/data/lab-results.json').then(r => r.json()).then(data => data.runs.filter(run => run.evidenceState === 'observed'))