electra-base-de-squad2
| Property | Value |
|---|---|
| Parameter Count | 111M |
| License | MIT |
| Language | German |
| Training Data | deQuAD2.0 (130k QA pairs) |
| Best Performance | 70.97% Exact Match, 76.18% F1 |
What is electra-base-de-squad2?
electra-base-de-squad2 is a German language question-answering model developed by Deutsche Telekom, built upon the ELECTRA architecture. It's specifically fine-tuned on deQuAD, a comprehensive German Question Answering dataset containing 130,000 training and 11,000 test QA pairs. This model represents a significant advancement in German language understanding and question-answering capabilities.
Implementation Details
The model is built on the electra-base-german-uncased architecture and trained using 8 V100 GPUs. It demonstrates superior performance compared to other German language models, particularly in exact match accuracy and F1 scores for both answerable and unanswerable questions.
- Base Architecture: ELECTRA German Uncased
- Training Dataset: deQuAD2.0 (~42MB training set)
- Evaluation Dataset: deQuAD2.0 test set (~4MB)
- Performance Metrics: 70.97% Exact Match, 76.18% F1 Score
Core Capabilities
- Advanced German text comprehension and analysis
- Accurate question-answering for both answerable and unanswerable queries
- Efficient context processing and answer extraction
- Easy integration with the Hugging Face transformers pipeline
Frequently Asked Questions
Q: What makes this model unique?
This model stands out for its superior performance in German question-answering tasks, outperforming other German language models like BERT variants. It achieves particularly strong results in both answerable (67.73% EM) and unanswerable (74.29% EM) questions.
Q: What are the recommended use cases?
The model is ideal for German language applications requiring precise question-answering capabilities, including automated customer service, information extraction from documents, and intelligent search systems. It's particularly effective when integrated into applications using the Hugging Face transformers pipeline.
