gemma3-examgen

Gemma3 ExamGen (Korean, XML)

TL;DR: A Gemma-3–based model fine-tuned to generate Korean university-level exam questions in strict XML (5 problems: 2 MCQ, 2 short-answer, 1 essay).

Outputs are in Korean.

---

Overview

Gemma3 ExamGen is a fine-tuned variant of Gemma-3 designed to generate Korean university exam questions in a strict XML structure.
It produces exactly five problems while enforcing the format and concept diversity.

---

Intended Use

• Primary : Generate Korean exam problems in XML.
• Output Language : Korean only.
• Not for : factual certification, grading, or unreviewed deployment.

---

Training Pipeline

• Base : google/gemma-3-4b-it
• Stages : SFT → DPO → GRPO
• Method : LoRA fine-tuning
• Data : PDF-crawled educational materials (private)
• Filtering : ensured XML validity and unique concepts.

---

Prompting Spec (Korean Prompt Template)

The model must always produce Korean outputs.
It strictly follows the XML schema and rules defined below.
When using this model, fill {KEYS} and {PHRS} placeholders with your own keywords and sentences extracted from context.

---

Prompt Template (in Korean)

다음의 규칙을 준수하여 대학교 시험 문제 5개를 XML 형식으로 생성하세요.

**응답 형식 (반드시 준수):**
<problems>
    <problem>
        <number>1</number>
        <type>객관식</type>
        <content>문제내용</content>
        <description>풀이과정</description>
        <answer>답</answer>
    </problem>
    <problem>
        <number>2</number>
        <type>객관식</type>
        <content>문제내용</content>
        <description>풀이과정</description>
        <answer>답</answer>
    </problem>
    
    <problem>
        <number>3</number>
        <type>단답형</type>
        <content>문제내용</content>
        <description>풀이과정</description>
        <answer>답</answer>
    </problem>
    <problem>
        <number>4</number>
        <type>단답형</type>
        <content>문제내용</content>
        <description>풀이과정</description>
        <answer>답</answer>
    </problem>
    
    <problem>
        <number>5</number>
        <type>주관식</type>
        <content>문제내용</content>
        <answer>답</answer>
    </problem>
</problems>

**절대 규칙 (위반 시 응답 무효):**
1. XML 태그 구조만 출력합니다. 다른 텍스트, 설명, 주석은 포함하지 않습니다.
2. 모든 내용은 CDATA 섹션 없이 일반 텍스트로 작성합니다.
3. 특수문자는 XML 엔티티로 작성합니다. (&lt;, &gt;, &amp;, &quot;, &apos;)

**문제 생성 규칙:**
- 총 5문제를 생성하며, 문제 유형은 다음 비율을 반드시 지킵니다: 객관식 2문제, 단답형 2문제, 주관식 1문제.
- 각 문제의 <type>은 위 응답 형식에서 이미 지정된 값을 그대로 사용합니다.
- 객관식 문제는 보기 기호를 ①, ②, ③, ④, ⑤ 형식으로 작성합니다.
- 모든 문제는 서로 다른 주요 개념을 사용해야 하며, 동일 개념이나 동일 인물, 동일 사건을 다른 문제에서 재사용하지 않습니다.
- 풀이과정과 답을 구체적으로 작성합니다.
- 문제 내용에 따옴표, 수식, 특수문자 등을 자유롭게 사용할 수 있습니다.
- 문제는 난이도와 표현 방식을 다양하게 구성합니다.

**중요한 키워드:**
{KEYS}
**중요한 문장들:**
{PHRS}

---

Example Usage

from transformers import AutoProcessor, AutoModelForImageTextToText
import torch

model_id = "yongjin-KIM/gemma3-examgen"
model = AutoModelForImageTextToText.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="auto")
processor = AutoProcessor.from_pretrained(model_id)
tok = processor.tokenizer

prompt = """<Insert the Korean prompt template here and replace {KEYS} and {PHRS}>"""

inputs = tok(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=2000,
    temperature=0.7,
    top_p=0.9,
    do_sample=True,
)
print(tok.decode(outputs[0], skip_special_tokens=True))

---

Output Format Guarantees

• Always produces well-formed XML.
• Exactly 5 <problem> blocks.
• Escapes all special characters (<, >, &, ", ').
• Fixed type order:
  객관식, 객관식, 단답형, 단답형, 주관식.

---

Limitations

• May occasionally omit <description> fields or produce overlong answers.
• Factual correctness is not guaranteed.
• Designed for Korean text only; English prompts are not supported.
• Contextual consistency may vary depending on {KEYS}/{PHRS} quality.

---

Ethical Considerations

• Intended for educational and research use only.
• Should not be used for unsupervised or high-stakes exam generation.
• All generated content should be reviewed by a human instructor before use.

---

Model Details

• Base Model: google/gemma-3-4b-it
• Architecture: Decoder-only transformer
• Fine-tuning Method: LoRA (r=8, α=32)
• Training Framework: PEFT + TRL
• Training Hardware: 2 × A100 (80GB)
• Training Duration: ~48 hours
• Stages: SFT → DPO → GRPO

---

License

• Model: CC-BY-NC-4.0
• Base Model: Gemma-3 (Google)
• Dataset: Private (PDF-crawled educational material)
• Intended Use: Research / Non-commercial

---

Maintainer

Author: Yongjin Kim
Hugging Face: @yongjin-KIM

---