Bespoke-MiniCheck-7B
| Property | Value |
|---|---|
| Parameter Count | 7.74B |
| Model Type | Text Classification |
| Architecture | InternLM2-based |
| License | CC BY-NC 4.0 |
| Paper | MiniCheck Paper |
What is Bespoke-MiniCheck-7B?
Bespoke-MiniCheck-7B is a state-of-the-art fact-checking model developed by Bespoke Labs. It's designed to determine whether a given sentence is supported by a reference document, outputting a binary classification (0 or 1). The model is built on the internlm2_5-7b-chat architecture and has been fine-tuned on a carefully curated dataset of 35K examples.
Implementation Details
The model is trained on a combination of 21K ANLI examples and 14K synthetically-generated examples, created using Meta's Llama-3.1-405B-Instruct. The synthetic data includes both "claim-to-document" and "doc-to-claim" examples, with sophisticated curation techniques to ensure high quality.
- Supports input documents up to 32K tokens
- Implements automatic prefix caching for improved performance
- Achieves throughput of >500 docs/min with vLLM optimization
- Uses BF16 tensor type for efficient computation
Core Capabilities
- Binary fact-checking classification
- Handles multi-sentence claims through sentence breakdown
- Efficient document processing with configurable chunk sizes
- State-of-the-art performance on LLM-AggreFact benchmark
- Supports batch processing for multiple documents and claims
Frequently Asked Questions
Q: What makes this model unique?
This model achieves SOTA performance in fact-checking despite its relatively small size, thanks to high-quality data curation and efficient architecture design. It's particularly notable for its combination of accuracy and processing speed.
Q: What are the recommended use cases?
The model is ideal for document-based fact verification, content validation, and automated fact-checking systems. It's particularly useful when processing large volumes of claims against reference documents, with practical applications in content moderation, research verification, and information validation.
