kez-lab/gemma-2-2b-quiz-korean
kez-lab/gemma-2-2b-quiz-korean is a specialized 2.6 billion parameter language model fine-tuned from Google Gemma-2-2B by KEZ Lab. It is engineered for source-grounded Multiple Choice Question (MCQ) generation from Korean text, strictly restricting question synthesis to facts explicitly verified within the provided source. The model utilizes internal Chain-of-Thought reasoning to generate pedagogically meaningful 4-choice questions with hard-negative distractors and exact evidence citations, achieving zero hallucination. It is optimized for on-device edge deployment on Android via Google MediaPipe Tasks GenAI and LiteRT.
Loading preview...
Overview
kez-lab/gemma-2-2b-quiz-korean is a specialized 2.6 billion parameter language model, fine-tuned by KEZ Lab from Google Gemma-2-2B. Its core function is to generate source-grounded Multiple Choice Questions (MCQs) directly from Korean blog posts, technical documentation, and long-form articles. A key differentiator is its strict adherence to factual grounding, ensuring that all generated questions are based solely on information explicitly present in the source text, thereby eliminating hallucinations.
Key Capabilities
- Internal Chain-of-Thought (
<thought>) Reasoning: Employs an autoregressive reasoning phase to plan evidence, target concepts, correct answers, and distractors before generating the final JSON output. - Zero-Hallucination Evidence Grounding: Each question includes a verbatim
evidencequote from the source, guaranteeing factual accuracy and auditability. - Equi-Distributed Answer Permutation: Eliminates position bias by uniformly distributing correct answers across options (A, B, C, D).
- Clean Korean Tokenization: Leverages Gemma-2's vocabulary for natural Korean syntax without CJK token leakage.
- On-Device Edge Deployment: Optimized for local inference on Android devices using Google MediaPipe Tasks GenAI and LiteRT, enabling zero-cloud dependency.
Performance & Benchmarks
The model achieves an Overall QuizScore of 0.985 / 1.000, demonstrating significant improvements over a prompted Base Gemma-2-2B. Notably, it scores 1.000 (100%) on Groundedness (Factuality), confirming zero hallucination, and 1.000 on Distractor Plausibility and Language Quality. This performance is evaluated against a diverse test suite of 100 document-level articles across 10 domains.
When to Use This Model
This model is ideal for applications requiring highly accurate, fact-checked, and pedagogically sound Korean MCQ generation from provided text. Its on-device optimization makes it suitable for mobile applications where offline functionality and low latency are critical, such as educational apps, content review tools, or knowledge assessment systems.