zebra-retriever-e5-base-v2
| Property | Value |
|---|---|
| Parameter Count | 109M |
| Base Model | intfloat/e5-base-v2 |
| License | CC BY-NC-SA 4.0 |
| Paper | arXiv:2410.05077 |
What is zebra-retriever-e5-base-v2?
zebra-retriever-e5-base-v2 is a specialized retrieval model designed as part of the ZEBRA (Zero-Shot Example-Based Retrieval Augmentation) framework for commonsense question answering. Built on the E5-base-v2 architecture, this model serves as the crucial first step in a three-stage pipeline that enhances the performance of language models in commonsense reasoning tasks.
Implementation Details
The model operates within the ZEBRA framework, which consists of three main stages: example retrieval, knowledge generation, and informed reasoning. This particular model handles the example retrieval phase, where it identifies relevant question-knowledge pairs from a large collection to support the answering process.
- Built on the E5-base-v2 architecture with 109M parameters
- Optimized for retrieval tasks in commonsense question answering
- Supports zero-shot capabilities without requiring task-specific training
- Integrates seamlessly with the ZEBRA pipeline
Core Capabilities
- Efficient retrieval of relevant example pairs
- Support for multiple question answering datasets including CSQA, ARC-C, PIQA
- Integration with various language models including Mistral-7B and Llama-3
- Demonstrated improvement in accuracy across multiple benchmarks
Frequently Asked Questions
Q: What makes this model unique?
This model is specifically designed for zero-shot retrieval in commonsense QA tasks, offering significant improvements in accuracy when used within the ZEBRA framework. It has shown consistent performance improvements across multiple benchmarks, with accuracy gains of up to 4-5 percentage points.
Q: What are the recommended use cases?
The model is ideal for commonsense question answering applications, particularly when integrated with larger language models. It's especially effective for tasks requiring contextual understanding and reasoning, such as educational applications, AI assistants, and automated question answering systems.
