KRLabsOrg/lettucedect-v2-qwen-2b
KRLabsOrg/lettucedect-v2-qwen-2b is a 2 billion parameter generative hallucination detector based on the Qwen3.5-2B architecture, developed by KRLabsOrg. This instruction-tuned model excels at localizing and typing hallucinated spans across diverse content, including code, tool output, and prose, in a single pass. It is specifically optimized for identifying unsupported information in RAG and coding-agent settings, outperforming larger models and specialized detectors in code-agent hallucination detection.
Loading preview...
Overview of lettucedect-v2-qwen-2b
lettucedect-v2-qwen-2b is a 2 billion parameter generative hallucination detector developed by KRLabsOrg, built on the Qwen3.5-2B base model. Unlike traditional token classifiers, this instruction-tuned model identifies and categorizes hallucinated spans within an answer by outputting structured JSON. It is uniquely trained on a unified benchmark encompassing code (SWE-bench-derived agent traces), prose (RAGTruth, PsiloQA, synthetic sources), and supports 14 languages, making it a versatile solution for diverse applications.
Key Capabilities and Differentiators
- Unified Hallucination Detection: Localizes and types hallucinated spans across code, tool output, and prose in a single pass, returning a JSON object with
text,category, andsubcategoryfor each unsupported span. - Superior Code-Agent Performance: Achieves a span-F1 of 0.602 and example-F1 of 0.835 on code-agent answers, significantly outperforming larger 120B and 550B judges, and other off-the-shelf detectors that tend to over-flag generated code.
- Competitive Prose Benchmarks: Performs strongly on established prose benchmarks like RAGTruth (example-F1 0.818) and PsiloQA (IoU 0.724 in English, 0.689 across 14 languages), often matching or exceeding specialized methods.
- Multilingual Support: Trained on PsiloQA, enabling robust hallucination detection across 14 languages.
- Generative Output: Provides a structured JSON output, which can also include per-span explanations, simplifying integration into downstream systems.
Ideal Use Cases
- RAG System Evaluation: Accurately identifies unsupported or contradictory information in responses generated by Retrieval-Augmented Generation systems.
- Code Agent Verification: Essential for validating the factual correctness and grounding of code generated by AI agents, where other detectors struggle.
- Multilingual Content Moderation: Useful for detecting hallucinations in generated text across various languages.
- Automated Content Quality Assurance: Integrates into pipelines to automatically flag and categorize hallucinated content, improving the reliability of AI-generated outputs.