WHAT IT SHOWS
The row-level evidence behind the aggregate score.
Every name and composition shown here comes from the link-free public export. Filters are encoded in the URL so a field note can point to the exact evidence it discusses.
FIELD NOTE 1 / BENCHMARK EXPLORER
Filter every frozen query by task, noise, correctness, or outcome. Open any row to inspect the candidate set and the model’s raw answer.
Research artifact. Not clinical guidance.
603 frozen rows
URL updates as filters change. Product links and IDs are not included in this export.
| Row | Input | Task / noise | Typesense | MedGemma | Outcome |
|---|
WHAT IT SHOWS
Every name and composition shown here comes from the link-free public export. Filters are encoded in the URL so a field note can point to the exact evidence it discusses.
WHAT IT DOES NOT SHOW
The explorer contains only the 603-row frozen benchmark. The held-out set contributes only its aggregate V0/V2 result to field note 1.