baselines/README.md

# Simple Baselines

This folder contains lightweight reference baselines for the packaged benchmark.

These scripts are designed for quick sanity checks and reproducible loading of the shared data package. They are not the paper's strongest models.

## Included scripts

- `property_prediction_baseline.py`: majority-class baseline for classification targets plus mean-RGB baseline for `color_rgb`.
- `image_generation_baseline.py`: nearest-neighbor retrieval baseline that matches each test image to the closest train sample in RGB space.

## Why these baselines exist

- They verify that the packaged JSON structure is directly consumable.
- They give downstream users a minimal executable reference without requiring the full project training stack.
- They produce simple metrics that are easy to compare after custom loaders or more advanced methods are added.

## Commands

Run the property prediction baseline:

```bash
python huggingface/baselines/property_prediction_baseline.py
```

Run the image generation retrieval baseline:

```bash
python huggingface/baselines/image_generation_baseline.py
```

Save image-generation retrievals for inspection:

```bash
python huggingface/baselines/image_generation_baseline.py --save-retrievals out/image_generation_retrievals.jsonl
```

## Output

- Both scripts print a JSON summary to stdout by default.
- `--output` writes the summary to a JSON file.
- `image_generation_baseline.py --save-retrievals ...` writes one JSON object per test sample with the retrieved train sample and image path.

## Notes

- No extra Python packages are required.
- The property baseline evaluates labeled test subsets only for `transparency` and `surface`.
- The image-generation baseline is retrieval-based, not a generative model. It is included as a minimal conditioning-aware reference for the packaged benchmark.