question-generator-model-qwen2.5-1.5b

Egyptian Ministry English Exam Question Generator (Qwen2.5-1.5B)

This model is a fine-tuned version of Qwen2.5-1.5B-Instruct designed to generate English multiple-choice questions that closely resemble the official examinations of the Egyptian Ministry of Education (General Secondary Stage).

Unlike generic question generation models, this model was trained specifically to reproduce the wording, structure, difficulty, distractor quality, and examination style commonly found in Egyptian Ministry exams.

---

Overview

The model receives curriculum information (Unit, Topic, Grammar Focus, Target Skills, Life Skills, Core Values) together with a user request and generates Ministry-style English MCQs in structured JSON format.

The objective is to assess understanding rather than memorization, while preserving the writing style used in official examinations.

---

Base Model

• Model: Qwen/Qwen2.5-1.5B-Instruct
• Fine-tuning: LoRA
• Framework: LLaMA-Factory
• Training Type: Supervised Fine-Tuning (SFT)
• Deployment Model: Merged LoRA + Base Model

---

Dataset

The model was trained using two complementary datasets.

1. Official Egyptian Ministry Dataset

A manually collected dataset extracted from previous official English examinations.

Each sample was converted into an instruction-following format including:

• curriculum metadata
• grammar focus
• target skills
• ministry-style output

---

2. Synthetic Dataset

To improve generalization, an additional synthetic dataset was generated using GLM-5.2.

The synthetic data follows the same curriculum structure and examination philosophy while introducing diverse contexts and wording.

This significantly increased dataset diversity without changing the Ministry examination style.

---

Capabilities

The model can generate:

• Grammar Questions
• Vocabulary Questions
• Reading Questions
• Writing Questions
• Ministry-style Distractors
• JSON Responses
• Multiple Questions in a Single Request

Each generated question contains:

• statement
• correct_answer
• plausible_distractors
• explanation

---

Recommended Prompt Structure

Although the model understands simple prompts, it performs significantly better when curriculum information is supplied.

Recommended prompt:

• User request
• Unit information
• Topic
• Grammar Focus
• Target Skills
• Life Skills
• Core Values
• JSON Output Schema

---

Example Prompt

Generate 5 Ministry-style grammar MCQs.

Unit: 11

Topic: Literature

Grammar Focus: Comparative and Superlative Adjectives

Return JSON only.

---

Loading the Model

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "tokhey/question-generator-model-qwen2.5-1.5b"

tokenizer = AutoTokenizer.from_pretrained(model_id)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype="auto"
)

---

Example Inference

messages = [
    {
        "role":"user",
        "content":"""
Generate 3 Ministry-style grammar MCQs.

Unit: 8

Topic:
A Journey Through Time

Grammar Focus:
Active and Passive Voice

Return JSON only.
"""
    }
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

inputs = tokenizer(text, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=1024,
    do_sample=False
)

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

---

Inference Tips

For deterministic evaluation:

do_sample=False

For more diverse generations:

do_sample=True
temperature=0.7
top_p=0.9

---

Intended Applications

• Educational Platforms
• AI Tutors
• Question Generation Systems
• Ministry-style Practice Exams
• Curriculum Assessment Systems
• Intelligent Learning Platforms

---

Limitations

• The model is optimized for Egyptian Ministry English examinations.
• Best performance is achieved when curriculum information is provided.
• Output quality depends on prompt specificity.

---

Acknowledgements

This model was developed as part of an undergraduate graduation project focused on automatic generation of Egyptian Ministry-style English examination questions using Large Language Models.

Base Model: Qwen Team

Fine-tuning Framework: LLaMA-Factory

Synthetic Data Generation: GLM-5.2

---