README.md

---
license: mit
task_categories:
- visual-question-answering
tags:
- medical-imaging
- vision-language-models
- ct
- hallucination
- benchmark
pretty_name: Priors Over Pixels — Result Data
---

# Priors Over Pixels — Result Data

Per-model result JSONs for the MICCAI 2026 **SAFER** workshop paper
*"Priors Over Pixels: Present-Bias in Organ-Presence Grounding for Medical VLMs."*

These are the raw model verdicts behind every table in the paper — a POPE-style
organ-presence probe on BTCV abdominal CT, testing whether medical VLMs ground
their answers in pixels or recite anatomical priors.

- **Code / reproduction:** https://github.com/MR-Nazarov/priors-over-pixels
- **Models probed:** MedGemma-4B/27B, Gemma3-4B/27B (base ablation), Qwen2.5-VL-7B, LLaVA-Med

## Files

| file | contents |
|---|---|
| `pope_results.json` | MedGemma-27B (main run) |
| `pope_results_medgemma4b.json` | MedGemma-4B |
| `pope_results_qwen.json` | Qwen2.5-VL-7B |
| `pope_results_llavamed.json` | LLaVA-Med |
| `gemma3_4b_pope.json` / `gemma3_27b_pope.json` | base-Gemma3 ablation |
| `pope_summary_allmodels.json`, `pope_abstention.json`, `adv_zdist_*.json`, `pope_noimg_summary.json` | derived summaries |

## Schema

Each file is `{"records": [...]}`; a record is one `(sample_id, organ, format)`
verdict:

```json
{
  "sample_id": "img0023_z070_prior_consistent_pancreas",
  "organ": "gallbladder",
  "neg_strategy": "adversarial",
  "ground_truth_present": false,
  "parsed_present": true,
  "parse_fail": false,
  "format": "freetext",
  "raw": "gallbladder, left kidney, pancreas"
}
```

## Usage

```
hf download Lexer1/priors-over-pixels-data --repo-type dataset --local-dir .
# then run the analysis scripts from the code repo
```

## License

MIT. Contains model outputs and derived statistics over public BTCV case
identifiers — no patient data. BTCV itself is not redistributed here.