---
license: mit
language:
- en
tags:
- glaze
- benchmark
pretty_name: GlazyBench
size_categories:
- 10K<n<100K
---
# GLAZE Benchmark for Hugging Face
[![PWC](https://img.shields.io/badge/%F0%9F%93%8E%20arXiv-Paper-red)](https://arxiv.org/abs/2605.06641)
[![PWC](https://img.shields.io/badge/GitHub-GlazyBench-black)](https://github.com/ziazhai/GlazyBench)

This folder is a publication-ready export of the benchmark assets that are currently treated as canonical in this repository.

It is organized into two benchmark tracks:

- `property_prediction/`: fixed train/test split for glaze property prediction.
- `image_generation/`: fixed train/test split for glaze image generation with paired metadata and conditioning signals.

It also contains an optional source-information layer for users who want to extract additional fields themselves:

- `source_records/`: one merged JSON file per packaged sample, plus filtered HTML-derived metadata.
- `raw_html/`: raw Glazy HTML pages for the packaged sample IDs only.
- `tools/`: small utilities that operate directly on the exported Hugging Face package.
- `baselines/`: minimal executable reference baselines for both benchmark tracks.

## Included benchmark assets

### 1. Property prediction benchmark

Canonical source: `data/`

Tasks:

- `transparency`: 4-class classification
  - Labels: `Opaque`, `Semi-opaque`, `Translucent`, `Transparent`
- `surface`: 9-class classification
  - Labels: `Glossy`, `Semi-glossy`, `Satin`, `Satin-matte`, `Matte`, `Semi-matte`, `Smooth Matte`, `Dry Matte`, `Stony Matte`
- `color_family`: 9-class classification
  - Labels: `Black`, `Blue`, `Gray`, `Green`, `Orange`, `Purple`, `Red`, `White`, `Yellow`
- `color_rgb`: RGB regression target derived from the same canonical target file

Current canonical split sizes:

- Train: 16,781 examples
- Test: 4,903 examples

Per-task labeled coverage in the current canonical files:

- Train transparency: 9,023
- Test transparency: 3,322
- Train surface: 9,378
- Test surface: 3,730
- Train color family: 16,781
- Test color family: 4,903

### 2. Image generation benchmark

Canonical source: `image_gen/`

This track contains images plus structured metadata for conditional image generation.

Current canonical split sizes:

- Train: 4,490 examples
- Test: 443 examples

Per-task labeled coverage in the current canonical files:

- Train transparency: 2,968
- Test transparency: 344
- Train surface: 3,381
- Test surface: 331
- Train color family: 4,483
- Test color family: 443

## Intentionally excluded from this export

These files exist in the repository but are not part of the public benchmark package because they are historical backups, intermediate variants, or analysis-side artifacts:

- `data/raw_data/`
- `data/train/targets_filtered_by_models.json`
- `data/train/targets_filtered_voting.json`
- `data/sample_types_train.csv`
- `data/sample_type_report.md`
- training logs, model outputs, checkpoints, and analysis-only files

The rule used here is simple: if a file is not part of the canonical benchmark split consumed by current benchmark code or benchmark-facing documentation, it is excluded.

## Optional source-information layer

Some fields in the canonical metadata are intentionally lightweight. In particular, page-derived attributes such as title, author, author URL, and long-form description are richer in the original HTML pages than in the benchmark-facing `metadata.json` files.

To support downstream custom extraction without changing the canonical splits, the package may include:

- `source_records/by_id/<id>.json`: merged per-sample record assembled from packaged recipe, target, metadata, and parsed HTML metadata.
- `source_records/html_metadata.json`: filtered HTML-derived metadata for packaged sample IDs.
- `raw_html/<id>.html`: original raw HTML page for the packaged sample ID.
- `tools/read_source_record.py`: convenience CLI for reading sample records and raw HTML from the exported package itself.

These supplemental files are aligned to packaged sample IDs and are meant for provenance and user-defined parsing, not as the primary benchmark interface.

## Minimal baselines included

To make the package directly usable after download, this export also includes dependency-light baseline scripts:

- `baselines/property_prediction_baseline.py`: majority-class baseline for `transparency`, `surface`, and `color_family`, plus a mean-RGB baseline for `color_rgb`.
- `baselines/image_generation_baseline.py`: nearest-train-sample retrieval baseline using RGB distance.

Run them from the repository root with:

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

See `baselines/README.md` for output details and optional flags.

## Publication note

This folder organizes the benchmark for sharing, but it does not assert a license on behalf of the original data sources. Before uploading to Hugging Face, confirm that the Glazy-derived metadata and images are cleared for redistribution under your intended release terms.

## Recommended Hugging Face presentation

If you publish this as a dataset repository, keep the two subfolders as two benchmark configurations inside one dataset card:

- `property_prediction`
- `image_generation`

This makes the public package match the benchmark structure already used in this repository.