---
title: README
emoji: 📚
colorFrom: pink
colorTo: green
sdk: static
pinned: false
---

**Welcome to Boldt!**

**Boldt** is a family of German language models developed by the **Chair of Machine Learning @ Humboldt-Universität zu Berlin**. This organization hosts our **models, datasets, and research artifacts** related to the Boldt project.

Feel free to explore, download, and experiment with our latest releases! 🚀

## 🌟 The Boldt Model Family

Our models are trained on our German *Dense-Core* subset of FineWeb-2, utilizing a multi-epoch training recipe on high-quality data. 

| Model | Parameters | Context Window | Description |
| :--- | :--- | :--- | :--- |
| [**Boldt-DC-350M**](https://huggingface.co/Boldt/Boldt-DC-350M) | 350M | 2048 | Ultra-lightweight base model for constrained environments. |
| [**Boldt-DC-1B**](https://huggingface.co/Boldt/Boldt-DC-1B) | 1B | 2048 | Highly optimized 1B base model with top-tier German performance. |
| [**Boldt-1B**](https://huggingface.co/Boldt/Boldt-1B) | 1B | 4096 | Extended context and augmented with 6B tokens of high-quality German news data. |
| [**Boldt-1B-IT-Preview**](https://huggingface.co/Boldt/Boldt-1B-IT-Preview) | 1B | 4096 | Experimental instruction-tuned model. |

## 📊 Comparison

Boldt-1B compares favorably on German LLM benchmarks against other similarly-sized models: 

![Boldt-1B Performance Comparison](https://huggingface.co/Boldt/Boldt-1B/resolve/main/boldt_1b_evaluation.png)

It is even competitive with many larger (2B parameter) models. See our paper for the full evaluation.

## 📖 Research & Artifacts
*   **Paper:** [Repetition over Diversity: High-Signal Data Filtering for Sample-Efficient German Language Modeling (arXiv 2026)](https://arxiv.org/abs/2604.28075)
*   **Evaluation Suite:** [Modernized German Benchmarks](https://huggingface.co/collections/Boldt/german-llm-benchmarks)