Egyptian MCQ Generator (Qwen2.5-7B)

This repository contains a fine-tuned version of Qwen/Qwen2.5-7B-Instruct for generating English multiple-choice questions in the style of the Egyptian Ministry of Education General Secondary examinations. The model was fine-tuned using LoRA and then merged into the base model for standalone inference.

Model Details

Base Model

• Qwen/Qwen2.5-7B-Instruct

Fine-tuning Method

• LoRA (PEFT)
• Merged with the base model

Task

• English Multiple-Choice Question Generation
• Grammar Question Generation
• Vocabulary Question Generation
• Curriculum-aware Question Generation

Dataset

The training dataset consists of:

• Authentic Egyptian Ministry examination questions.
• Curriculum metadata extracted from the official syllabus.
• Synthetic grammar questions.
• Synthetic vocabulary questions.

Each training sample follows the chat format:

• System Prompt
• User Prompt
• Assistant JSON Response

The model was trained to produce structured JSON outputs rather than plain text.

Supported Prompt Styles

The model has been trained on multiple prompt formulations, including:

• Full curriculum prompts
• Grammar-only prompts
• Vocabulary-only prompts
• Topic-based prompts
• Unit-based prompts
• Unit + Grammar prompts
• Topic + Grammar prompts

This improves prompt robustness and instruction following.

Output Format

The model returns JSON only. Example:

{
  "questions": [
    {
      "statement": "...",
      "correct_answer": "...",
      "plausible_distractors": [
        "...",
        "...",
        "..."
      ],
      "explanation": "..."
    }
  ]
}

Example Prompt

Generate 10 grammar questions.
Grammar Focus:
Past Perfect / Past Perfect Continuous
Return only valid JSON.

or

Generate 10 questions.

Topic:
At the Airport

Return only valid JSON.

Intended Use

This model is intended for:

• Educational applications
• Automatic question generation
• Curriculum-aligned assessment
• Research in educational NLP
• English language learning systems

Limitations

This is an experimental research model. Although it has been fine-tuned on Ministry-style data, generated questions should be reviewed by educators before use in real examinations.

Acknowledgements

• Qwen Team
• Hugging Face
• Unsloth
• PEFT
• TRL

License

Apache-2.0