Dataset Card for Konkur1404

Dataset Description

This dataset contains questions from the Konkur (Iranian University Entrance Exam) for the year 1404. It is designed for evaluating models on Persian multiple-choice questions across various subjects.

Dataset Summary

• Total Examples: 2137
• Splits: train
• Languages: Persian (fa)

Dataset Structure

Data Instances

An example from the dataset looks like this:

{
  "id": "ensani_nobat1_1",
  "exam_name": "ensani_nobat1",
  "question": "اگر شعاع دایره شکل زیر برابر $x = \\frac{1}{\\sqrt{2\\pi}}$ و مجموع مساحتهای دو شکل برابر ۱۶ باشد، محیط دایره کدام است؟",
  "choices": [
    "$\\sqrt{\\pi}$",
    "$2\\sqrt{\\pi}$",
    "$3\\sqrt{\\pi}$",
    "$4\\sqrt{\\pi}$"
  ],
  "answer_key": 4,
  "figure": "<Image: PNG, (437, 231)>"
}

Data Fields

The dataset contains the following fields:

• id (string): Description of id.
• exam_name (string): Description of exam_name.
• question (string): Description of question.
• choices (List(Value('string'))): Description of choices.
• answer_key (int32): Description of answer_key.
• figure (PIL.Image.Image): Description of figure.

Dataset Statistics

Split: train

• Count: 2137
• exam_name Distribution:
  • zaban_nobat1: 400
  • zaban_nobat2: 350
  • ensani_nobat1: 280
  • tajrobi_nobat1: 225
  • ensani_nobat2: 221
  • tajrobi_nobat2: 185
  • riazi_nobat1: 145
  • honar_nobat1: 126
  • riazi_nobat2: 105
  • honar_nobat2: 100
• answer_key Distribution:
  • 1.0: 551
  • 2.0: 539
  • 3.0: 538
  • 4.0: 508

Evaluation with OpenAI-Compatible API

• Deterministic settings (temperature=0) are recommended.
• Normalize Persian digits and English number words.
• Report both overall accuracy and per-exam accuracy.
• Use multimodal input for questions with figures if your model supports images.

Evaluation Script

import os
import io
import base64
import re
import csv
import time
from collections import defaultdict

from openai import OpenAI
from datasets import load_dataset
from tqdm import tqdm

API_KEY = os.getenv("OPENROUTER_API_KEY") or os.getenv("OPENAI_API_KEY", "your-api-key")
BASE_URL = os.getenv("OPENROUTER_BASE_URL", "https://openrouter.ai/api/v1")
MODEL_NAME = os.getenv("OPENROUTER_MODEL", "openai/gpt-5.2")
USE_IMAGES = True
EXAMS = ["ensani_nobat1", "ensani_nobat2"]

client = OpenAI(api_key=API_KEY, base_url=BASE_URL)

def format_prompt(example):
    prompt = f"Question: {example['question']}\n\n"
    for i, choice in enumerate(example['choices']):
        prompt += f"{i+1}. {choice}\n"
    prompt += "\nAnswer with the number of the correct choice (1, 2, 3, or 4) only."
    return prompt

def extract_answer(response_text):
    text = response_text.strip()
    for k, v in {"۱": "1", "۲": "2", "۳": "3", "۴": "4"}.items():
        text = text.replace(k, v)
    for k, v in {"one": "1", "two": "2", "three": "3", "four": "4"}.items():
        if re.search(rf"\b{k}\b", text, flags=re.IGNORECASE):
            text = v
            break
    for k, v in {"یک": "1", "يك": "1", "دو": "2", "سه": "3", "چهار": "4"}.items():
        if k in text:
            text = v
            break
    m = re.search(r"\b([1-4])\b", text)
    return int(m.group(1)) if m else None

def figure_to_base64(figure):
    if not figure:
        return None
    try:
        if hasattr(figure, "save"):
            buf = io.BytesIO()
            figure.save(buf, format="PNG")
            return base64.b64encode(buf.getvalue()).decode("utf-8")
        if isinstance(figure, str):
            path = figure
            if not os.path.isabs(path):
                path = os.path.join(os.getcwd(), path)
            from PIL import Image
            img = Image.open(path)
            buf = io.BytesIO()
            img.save(buf, format="PNG")
            return base64.b64encode(buf.getvalue()).decode("utf-8")
    except Exception:
        return None
    return None

def chat_with_retries(messages, max_retries=3):
    delay = 1.0
    for attempt in range(max_retries):
        try:
            return client.chat.completions.create(
                model=MODEL_NAME,
                messages=messages,
                temperature=0,
                max_tokens=10
            )
        except Exception:
            if attempt == max_retries - 1:
                raise
            time.sleep(delay)
            delay = min(8.0, delay * 2)

def evaluate():
    ds = load_dataset("mshojaei77/konkur1404", split="train")
    if EXAMS:
        ds = ds.filter(lambda x: x.get("exam_name") in EXAMS)

    totals = defaultdict(int)
    corrects = defaultdict(int)
    rows = []

    for example in tqdm(ds):

        prompt = format_prompt(example)
        messages = [{"role": "system", "content": "Answer only with 1, 2, 3, or 4."}]

        img_b64 = None
        if USE_IMAGES:
            img_b64 = figure_to_base64(example.get("figure"))

        if USE_IMAGES and img_b64:
            messages.append({
                "role": "user",
                "content": [
                    {"type": "text", "text": prompt},
                    {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{img_b64}"}}
                ]
            })
        else:
            messages.append({"role": "user", "content": prompt})

        pred = None
        error_msg = ""
        try:
            resp = chat_with_retries(messages)
            prediction_text = resp.choices[0].message.content.strip()
            pred = extract_answer(prediction_text)
        except Exception as e:
            error_msg = str(e)

        gt = int(example["answer_key"])
        exam = example.get("exam_name", "unknown")
        totals[exam] += 1
        ok = int(pred == gt)
        corrects[exam] += ok
        rows.append({"id": example.get("id"), "exam_name": exam, "predicted": pred, "ground_truth": gt, "correct": ok, "error": error_msg})
        if error_msg:
            print(f"Error on id={example.get('id')} exam={exam}: {error_msg}")

    total = sum(totals.values())
    correct = sum(corrects.values())
    if total:
        print(f"Accuracy: {100*correct/total:.2f}% ({correct}/{total})")
        for exam, t in totals.items():
            if t:
                print(f"- {exam}: {100*corrects[exam]/t:.2f}% ({corrects[exam]}/{t})")
    else:
        print("No examples evaluated.")

    if rows:
        with open("konkur1404_results.csv", "w", newline="", encoding="utf-8") as f:
            w = csv.DictWriter(f, fieldnames=["id","exam_name","predicted","ground_truth","correct","error"])
            w.writeheader()
            w.writerows(rows)
        print("Saved konkur1404_results.csv")

if __name__ == "__main__":
    evaluate()

Data Notes

• Choices are always 4 options; answer_key is 1–4 (1-based).
• Figures are PNGs referenced by relative paths; when loaded via HF Datasets, figure may be an image object.
• Text may include LaTeX-style math and Persian digits; normalize for robust parsing.

Ethics and Usage

• For evaluation and research use; respect exam policies and local regulations.
• Random baseline is 25% accuracy; report per-exam breakdown for interpretability.