Elicitation format changes zero-shot classification accuracy in a medical vision-language model: code and run records
Description
Code, configurations, prompts and sealed run records for a two-center evaluation of elicitation format on nine-class colorectal histopathology classification (PathMNIST, MedMNIST v2) with MedGemma-27b-it.
Holding the prompt fixed down to its per-class cue block and varying only the requested response format, asking for a probability over all nine classes rather than a ranked top-3 costs 14.8 accuracy points on validation and 13.4 at an external clinical center. An answer-suppression control locates the cost in the closing sentence forbidding commentary, reasoning and a chosen answer rather than in the requested cardinality.
The served endpoint is not reproducible by a third party, so every raw model response is released: each of the 123 run directories carries its configuration, that configuration's SHA-256, the patch manifest, the raw responses, per-patch token usage, run metadata and a checksum manifest over all of it. Every number in the accompanying article resolves to exactly one entry of the released append-only results log.
No model weights are redistributed.
Notes
Files
Files
(18.4 MB)
| Name | Size | |
|---|---|---|
|
md5:de0b564671ca6d739b41866502b0b04f
|
18.4 MB | Download |
Additional details
Related works
- Is derived from
- 10.5281/zenodo.22319249 (DOI)
- Is supplement to
- https://github.com/anayy09/elicitation-format-medvlm/tree/v2.0.0 (URL)
- References
- 10.5281/zenodo.1214456 (DOI)