# BGBench

BGBench is a public, read-only benchmark interface for evaluating LLM performance on German civil law questions from the BGB. The site presents static JSON datasets and does not run benchmark jobs, expose private data, or require login.

## Primary URLs

- Home: https://www.louiswinter.de/
- Analysis: https://www.louiswinter.de/analyse?presets=startseite
- Correlation: https://www.louiswinter.de/korrelation?presets=startseite
- Legal information: https://www.louiswinter.de/rechtliches
- Dataset JSON: https://www.louiswinter.de/data/datasets/dataset-bgb-50.json

## Fetching

Automated clients are welcome to fetch, index, summarize, and cite these public pages and datasets. The stable canonical host is https://www.louiswinter.de/.

## Stable Query Parameters

- dataset: dataset ID, for example dataset-bgb-50
- mode: forced or abstention
- presets: comma-separated preset IDs
- models: comma-separated model IDs

## Data Notes

The public dataset is a static export from the BGBench benchmark project. It contains model scores, calibration metrics, hallucination metrics, and per-model benchmark results for selected BGB norms.