felfri/dose-response-generated-images

---
license: apache-2.0
tags:
- diffusion
- text-to-image
- safety
- dose-response
- image-generation
size_categories:
- 10K<n<100K
---

# Dose-Response Generated Images

Generated images and safety annotations from the dose-response experiment studying how unsafe training data fraction affects text-to-image model output safety.

## Dataset Description

Each condition's model was used to generate 10,000 images from the same prompt set (1K safe + 9K unsafe prompts). All images were annotated for safety using 4 independent safety evaluators.

## Structure

```
images/
C0/images.tar # 10K images from C0 model (0% unsafe training data)
C1/images.tar # 10K images from C1 model (5% unsafe)
C3/images.tar # 10K images from C3 model (~1.21% unsafe, original)
C4/images.tar # 10K images from C4 model (~1.21% unsafe, 1M scale)
C5/images.tar # 10K images from C5 model (~9.6% unsafe, 1M scale)
C6/images.tar # 10K images from C6 model (~1.21% unsafe, 100K scale)
annotations/
dose_C0.parquet # Unified annotations from all 4 evaluators
dose_C1.parquet
dose_C3.parquet
dose_C4.parquet
dose_C5.parquet
dose_C6.parquet
```

## Annotation Format

Each parquet file is indexed by image ID and contains the following columns:

| Column | Type | Description |
|--------|------|-------------|
| `llavaguard_unsafe` | int | 0 = safe, 1 = unsafe (LlavaGuard-7B) |
| `llavaguard_category` | str | Safety category (e.g. "O3: Sexual Content", "NA: None applying") |
| `llamaguard3_unsafe` | int | 0 = safe, 1 = unsafe (LlamaGuard-3-11B-Vision) |
| `llamaguard3_category` | str | MLCommons category (e.g. "S12: Sexual Content", "NA") |
| `shieldgemma_unsafe` | int | 0 = safe, 1 = unsafe (ShieldGemma-2-4B) |
| `shieldgemma_category` | str | Highest-probability policy violated ("dangerous", "sexual", "violence", or "NA") |
| `shieldgemma_p_dangerous` | float | P(dangerous content) |
| `shieldgemma_p_sexual` | float | P(sexual content) |
| `shieldgemma_p_violence` | float | P(violent content) |
| `sd_safety_checker_unsafe` | int | 0 = safe, 1 = unsafe (SD Safety Checker) |

## Safety Evaluators

| Evaluator | Model | Categories |
|-----------|-------|------------|
| [LlavaGuard-7B](https://huggingface.co/AIML-TUDA/LlavaGuard-v1.2-7B-OV) | Vision-language safety model | O1–O9 (9 categories) |
| [LlamaGuard-3-11B-Vision](https://huggingface.co/meta-llama/Llama-Guard-3-11B-Vision) | Meta's multimodal safety model | S1–S14 (MLCommons taxonomy) |
| [ShieldGemma-2-4B](https://huggingface.co/google/shieldgemma-2-4b) | Google's safety classifier | dangerous, sexual, violence |
| SD Safety Checker | CompVis CLIP-based NSFW classifier | binary (safe/unsafe) |

## Generation Settings

- **Prompts**: 10,000 (1K safe + 9K unsafe across 9 categories)
- **Guidance scale**: 3.5
- **Inference steps**: 50
- **Seed**: 42
- **Resolution**: 512px

## Related Resources

- [diffusion_safety](https://github.com/felifri/diffusion_safety) — experiment code
- Model checkpoints: `felfri/dose-response-c0` through `felfri/dose-response-c6`