---
license: cc-by-4.0
task_categories:
- visual-question-answering
tags:
- vlm
- robustness
- answer-flip
- benchmark
pretty_name: FlipDir Answer-Flip Testbeds
---

# FlipDir — Answer-Flip Testbeds under Benign Visual Variations

Model-conditioned testbeds of **answer flips**: cases where a vision–language model
changes its final answer under a benign visual variation (exposure, white balance,
JPEG, blur, small rotation) that a human would treat as inconsequential. Companion
data for the FlipDir defense — code at **https://github.com/YeonsungJung/FlipDir**.

## What is (and isn't) here

Each testbed ships the **metadata + seeds + variation parameters + flip labels**, not
the rendered images. Because generation is deterministic (batch-1 greedy, fixed seed
chain), the varied images regenerate **byte-for-byte** from the original source image +
the stored seed/params using the released code. This keeps the release small and avoids
redistributing the source datasets' images.

Regenerate images:
```bash
# in the code repo, with the same seed
bash run_ours.sh <qwen8b|gemma12b> <attack> gen
```

## Layout

```
paper/<model>/<testbed>/<split>/meta/*.json     # settings used in the paper
etc/<model>/<testbed>/<split>/meta/*.json        # extra intensity/seed variants
  <model>   = qwen3vl-8b | gemma-3-12b
  <testbed> = exposure_shift | wb_shift | jpeg_double            (M3CoT, science)
              exposure_shift_robo | brightness_shift_robo | motion_shake_robo (Robo2VLM)
              window_level | medical_blur | detector_rotation    (GMAI-MMBench, medical)
  <split>   = train | test
```

A flip label is defined relative to the model's own clean-image answer, so every testbed
is built **per model**. Current release: Gemma-3-12B all 9 testbeds; Qwen3-VL-8B 7 of 9
(robotics `exposure_shift_robo` / `brightness_shift_robo` are being regenerated).

## Metadata schema (per parent question)

`id, split, category_rank, attack, text_attack, image_attack_intensity, model, seed, gt,
prompt, orig_img_path, orig_raw, orig_norm, orig_correct, raw_counts, flips, nonflips,
limits, gen_cfg` — `orig_norm` is the clean-image answer, `flips`/`nonflips` list the
varied children and their (raw, normalized) outputs, `seed`+`image_attack_intensity`
reproduce the variation, `orig_img_path` points into the source dataset.

## Source datasets (obtain images from the originals)

- **M3CoT** — science reasoning · https://huggingface.co/datasets/LightChen2333/M3CoT
- **Robo2VLM** — robot-scene VQA · https://huggingface.co/datasets/keplerc/Robo2VLM
- **GMAI-MMBench** — medical VQA · https://huggingface.co/datasets/OpenGVLab/GMAI-MMBench

Please follow each source dataset's license for the underlying images.

## Citation

FlipDir (preprint). Code and paper: https://github.com/YeonsungJung/FlipDir