Gemma-2-2B: Source-Grounded Korean MCQ Generator

Base Model: Google Gemma-2-2B • Benchmark Score: 0.985 / 1.000 • 100% Grounded • Android LiteRT Ready

gemma-2-2b-quiz-korean is a specialized lightweight language model fine-tuned from Google Gemma-2-2B. It is engineered specifically for source-grounded Multiple Choice Question (MCQ) generation directly from Korean blog posts, technical documentation, and long-form articles.

The model strictly restricts question synthesis to facts explicitly verified within the provided source text, eliminating hallucinations while generating pedagogically meaningful 4-choice questions with hard-negative distractors and exact evidence citations.

Key Capabilities

1. Internal Chain-of-Thought (<thought>) Reasoning: Unlike standard single-pass generators, the model utilizes an internal autoregressive reasoning phase before producing the final JSON payload. It systematically plans the evidence sentence, target concept, unambiguous correct answer, and three same-category distractors.
2. Zero-Hallucination Evidence Grounding: Every generated question includes a verbatim evidence quote from the source text, ensuring strict factual alignment and auditability.
3. Equi-Distributed Answer Permutation: The model eliminates position bias (such as index-0 bias) through uniform answer position training (25% distribution across A, B, C, D).
4. Clean 256k Multilingual Tokenization: Leveraging Gemma-2's vocabulary, the model generates natural Korean syntax without CJK token leakage or Hanja bleeding.
5. On-Device Edge Deployment: Optimized for local inference via Google MediaPipe Tasks GenAI and LiteRT (TensorFlow Lite), running on modern Android devices with zero cloud dependency.

Benchmark Evaluation

Evaluated against a held-out test suite of 100 document-level isolated articles across 10 domains (Android, Distributed Systems, AI/ML, Economics, Medicine, Natural Sciences, History, Philosophy, UX Design, and Architecture).

$$\text{QuizScore} = 0.30 \times \text{Groundedness} + 0.25 \times \text{Uniqueness} + 0.20 \times \text{DistractorQuality} + 0.15 \times \text{Importance} + 0.10 \times \text{LanguageQuality}$$

Output Schema

The model generates an internal <thought> trace followed by a structured JSON payload:

{
  "questions": [
    {
      "question": "Specific question grounded in the source text",
      "options": [
        "Option A (same semantic category)",
        "Option B (same semantic category)",
        "Option C (same semantic category)",
        "Option D (same semantic category)"
      ],
      "answer_index": 2,
      "explanation": "Clear educational explanation of why the answer is correct",
      "evidence": "Verbatim excerpt from the article proving the answer"
    }
  ]
}

Quickstart (Python / Transformers)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL_ID = "kez-lab/gemma-2-2b-quiz-korean"

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

article = """
Jetpack Compose에서 derivedStateOf를 사용하면 원본 상태의 잦은 변화 중에서
우리가 관심 있는 특정 조건(firstVisibleItemIndex > 0)이 변경되는 순간에만
다운스트림 Recomposition을 트리거하도록 캐싱 및 완충 역할을 합니다.
"""

messages = [
    {
        "role": "user",
        "content": f"주어진 글만을 근거로 핵심 개념을 분석하고 4지선다 객관식 문제를 생성하라.\n\n[ARTICLE]\n{article}"
    }
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=1024,
        temperature=0.01,
        repetition_penalty=1.1
    )

response = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
print(response)

Android On-Device SDK Integration

This model can be compiled into .task / .tflite format for on-device execution using Google MediaPipe Tasks GenAI:

// Android SDK Usage
val generator = LocalQuizGenerator.builder(context)
    .fromHuggingFace("kez-lab/gemma-2-2b-quiz-korean")
    .setMaxTokens(1024)
    .setTemperature(0.01f)
    .build()

viewModelScope.launch {
    val result = generator.generateQuiz(articleText)
    result.onSuccess { quizzes ->
        // Render verified quiz cards with evidence quotes
    }
}

Training Details

• Base Architecture: Google Gemma-2-2B (unsloth/gemma-2-2b-it, 2.6B parameters)
• Methodology: Parameter-Efficient Fine-Tuning (PEFT / LoRA) with Standalone Weight Merge
• LoRA Configuration: $r=16, \alpha=32$, targeting all linear projections (q, k, v, o, gate, up, down)
• Precision: bfloat16
• Loss Masking: Response-only loss computation on assistant reasoning and JSON tokens
• Dataset: Multi-domain curriculum of 1,000 verified CoT samples validated through strict Critic filtering

License & Attribution

• Base Model License: Gemma Terms of Use
• Developed by: KEZ Lab