| Property | Value |
|---|---|
| Parameter Count | 336M |
| Model Type | Sentence Transformer |
| Architecture | gBERT-large based with mean pooling |
| Dimension | 1024 |
What is German_Semantic_STS_V2?
German_Semantic_STS_V2 is a state-of-the-art German language model specifically designed for semantic similarity tasks. Built on gBERT-large architecture, it maps sentences and paragraphs to 1024-dimensional dense vector space, achieving an impressive 86.26% Spearman correlation on German STS benchmarks.
Implementation Details
The model uses a sentence-transformers framework with a two-component architecture: a transformer module with 512 max sequence length and a pooling layer implementing mean token pooling. It was trained using ContrastiveLoss with cosine distance metric and a margin of 0.5.
- Trained with AdamW optimizer (lr=5e-06)
- 4 epochs with 576 warmup steps
- Batch size of 4 with gradient clipping at 1.0
Core Capabilities
- Semantic text similarity scoring
- Clustering of German text
- Semantic search applications
- Cross-lingual tasks (particularly German-English)
- Domain and intent classification
Frequently Asked Questions
Q: What makes this model unique?
This model outperforms previous German language models on semantic similarity tasks, achieving higher Spearman correlation (86.26%) than established multilingual models like XLM-R and other German-specific models.
Q: What are the recommended use cases?
The model excels in semantic search, document clustering, and similarity analysis for German text. It's particularly effective for tasks requiring nuanced understanding of semantic relationships between texts.