KRLabsOrg/lettucedect-v2-qwen-2b

Overview: Generative Hallucination Span Detection

lettucedect-v2-qwen-2b is a 2.3 billion parameter instruction-tuned generative model from KRLabsOrg, built on Qwen3.5-2B. It functions as a unified hallucination detector, identifying and categorizing unsupported spans within generated text relative to a given context. Unlike token classifiers, this model outputs structured JSON, providing exact hallucinated spans with their category and subcategory in a single pass.

Key Capabilities & Differentiators

• Unified Span Detection: Localizes and types hallucinated spans across diverse content, including prose (RAGTruth, PsiloQA), code (SWE-bench-derived agent traces), and tool output.
• Multilingual Support: Trained on a unified benchmark covering English and 13 other languages via the PsiloQA dataset.
• Superior Code Hallucination Detection: 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 Performance: Matches or exceeds specialized methods on established prose benchmarks like RAGTruth (example-F1 0.818) and PsiloQA (IoU 0.724 in English).
• Structured Output: Returns a JSON object detailing each hallucinated span, its category (contradiction, fabricated_reference, unsupported_addition), and subcategory.

Ideal Use Cases

This model is particularly well-suited for developers and researchers needing a fast, small, and accurate solution for:

• RAG System Evaluation: Identifying and mitigating hallucinations in RAG-generated responses.
• Code Agent Verification: Detecting unsupported or fabricated elements in code generated by AI agents.
• Multilingual Content Moderation: Ensuring factual consistency across various languages.
• Automated Content Quality Assurance: Providing granular feedback on factual accuracy at the span level.