MiniCheck-RoBERTa-Large
| Property | Value |
|---|---|
| License | MIT |
| Research Paper | MiniCheck: Efficient Fact-Checking of LLMs on Grounding Documents |
| Primary Task | Text Classification (Fact Checking) |
| Language | English |
What is MiniCheck-RoBERTa-Large?
MiniCheck-RoBERTa-Large is a specialized fact-checking model designed to verify whether claims are supported by reference documents. Built on RoBERTa-Large architecture, it performs binary classification at the sentence level, determining if a given claim is supported (1) or unsupported (0) by the provided document. The model was fine-tuned on 14K synthetic data points, specifically structured for fact-checking tasks.
Implementation Details
The model builds upon the AlignScore RoBERTa-Large foundation and processes input in a document-claim pair format. It outputs both binary predictions and confidence scores, making it suitable for automated fact-verification workflows. The implementation achieves approximately 800 documents per minute processing speed, depending on hardware capabilities.
- Binary classification output (0 or 1)
- Confidence scores for predictions
- Efficient processing pipeline
- Easy integration through Python package
Core Capabilities
- Sentence-level fact verification against reference documents
- High accuracy on the LLM-AggreFact benchmark
- Batch processing of multiple document-claim pairs
- Real-time confidence scoring
Frequently Asked Questions
Q: What makes this model unique?
The model outperforms existing specialized fact-checkers of similar scale on the LLM-AggreFact benchmark, which includes data from 11 recent human-annotated datasets. It's specifically designed for efficient, accurate fact-checking without requiring human intervention or synthetic error injection.
Q: What are the recommended use cases?
The model is ideal for automated fact-checking systems, content verification pipelines, and research applications requiring document-grounded claim verification. It's particularly well-suited for validating LLM-generated content against source documents.
