LettuceDetect Base ModernBERT
| Property | Value |
|---|---|
| Organization | KRLabsOrg |
| Architecture | ModernBERT with extended context |
| Task | Hallucination Detection |
| Context Length | 8192 tokens |
| Language | English |
| Paper | arXiv:2502.17125 |
What is lettucedect-base-modernbert-en-v1?
LettuceDetect is a sophisticated transformer-based model specifically designed for hallucination detection in Retrieval-Augmented Generation (RAG) applications. Built on ModernBERT architecture, it excels at analyzing context-answer pairs to identify potential hallucinations in generated text. With its impressive 8192-token context support, it offers superior capability in processing extensive documents for accuracy verification.
Implementation Details
The model employs token classification architecture and has been trained on the RagTruth dataset. It performs token-level analysis to identify sections of answers that aren't supported by the provided context, aggregating these into meaningful spans for easy interpretation.
- Token-level hallucination detection with span aggregation
- Extended context support up to 8192 tokens
- Easy integration through Python API
- State-of-the-art performance metrics
Core Capabilities
- Accurate identification of hallucinated content in generated text
- Span-level prediction output for precise hallucination location
- Support for extensive context processing
- F1 score of 79.22% on RAGTruth test set
- Competitive performance against larger language models
Frequently Asked Questions
Q: What makes this model unique?
LettuceDetect stands out for its extended context support and competitive performance against larger models like fine-tuned LLAMA-2-13B, while maintaining efficient inference times. Its token-level analysis provides granular insight into hallucinated content.
Q: What are the recommended use cases?
The model is particularly suited for RAG applications where accuracy verification is crucial. It's ideal for content verification systems, automated fact-checking, and quality assurance in AI-generated content.
