metadata

license: apache-2.0
language:
  - en
tags:
  - mathematics
  - computer-science
  - cryptograpy
  - ctf
pretty_name: Dynamic Intelligence Assessment Dataset
configs:
  - config_name: K1
    data_files: DIA-Benchmark-k1.json
    type: json
    field: questions
  - config_name: K5
    data_files: DIA-Benchmark-k5.json
    type: json
    field: questions
  - config_name: K10
    data_files: DIA-Benchmark-k10.json
    type: json
    field: questions
  - config_name: K100
    data_files: DIA-Benchmark-k100.json
    type: json
    field: questions

Dynamic Intelligence Assessment Dataset

This dataset aims to test the problem-solving ability of LLMs with dynamically generated challenges that are difficult to guess.

Dataset Details

The DIA Benchmark Dataset is a benchmarking tool consisting of 150 dynamic question generators for the evaluation of the problem-solving capability of LLMs. It primarily focuses on CTF-style (Capture the Flag) challenges that require knowledge from the fields of mathematics, cryptography, cybersecurity, and computer science. The challenge generators were manually developed by industry experts and tested by multiple individuals to find errors and edge cases. The answers often consist of many characters and big numbers, making correct guessing highly unlikely. This repository contains the generated question and answer pairs that can be fed to AI models to assess the outputs. The repository contains various generated instances of one test to increase the accuracy of the measurements.

Curated by: Norbert Tihanyi, Tamas Bisztray, Richard A. Dubniczky, Rebeka Toth, Bertalan Borsos, Bilel Cherif, Ridhi Jain, Lajos Muzsai, Mohamed Amine Ferrag, Ryan Marinelli, Lucas C. Cordeiro, Merouane Debbah, Vasileios Mavroeidis, and Audun Josang
Language: English
License: AL 2.0

Evaluation

We tested 25 state-of-the-art LLM models on the DIA dataset through API calls, and ChatGPT-4o manually through its chat interface to enable tool usage. The tests were generated and run in November 2024 on the k=5 dataset.

Sources

Repository: https://github.com/DIA-Bench/DIA-Bench
Paper: https://arxiv.org/abs/2410.15490

Uses

Testing the problem-solving ability of an LLM.
Comparing the problem-solving abilities of multiple LLMs.
Testing the response consistency of an LLM.
Testing the confidence of an LLM.

Dataset Structure

During generation, different sizes of the dataset can be created by modifying the k parameter, where k=5 means that one test generator will be used to generate 5 instances of that test type. This repository contains four variations of the pre-generated tests datiled in the following table:

File	K	Tests
DIA-bench-1.json	1	150
DIA-bench-5.json	5	750
DIA-bench-10.json	10	1,500
DIA-bench-100.json	100	15,000

Please note that depending on the tool or API used to run the tests, the runtime as well as the cost will differ significantly.

Each generated dataset is a JSON file, where the questions are stored in a list under the "questions" key. An example question looks as follows:

{
    "challenge": {
        "template_id": 1,
        "instance": 1,
        "level": "easy",
        "category": "Mathematics",
        "adversarial": false,
        "description": "You have been given the task of solving the following challenge.\n\n        #Instructions:\n              a. Carefully read the challenge description.\n              b. Determine the most effective approach to solve the challenge. This may include writing a program, utilizing external resources, employing specific tools, or any other suitable methods.\n              c. We are solely interested in the solution, not the method used to obtain it.\n              d. Confidence in your solution is extremely important. If your answer is correct, you will get +1 point. If incorrect, you will get -2 point. If you are unsure of the solution, or you can not solve it respond in this XML format:  <xml>I-DO-NOT-KNOW</xml> to receive 0 points without penalty for an incorrect answer.\n              e. Be sure to double-check your solution before answering and  make sure your answer follows the required XML format as specified in the challenge.\n\n\n        #Challenge:\n\n        Given a 128-bit even number, namely 170272422572648913815378456437965460739, find the next prime number.",
        "instructions": "Find the next prime number that comes after the specified number. Do NOT include any explanation or additional text in the response. ALWAYS return the answer in this XML format: '<xml>prime_number</xml>'."
    },
    "solution": {
        "challenge_solution": "<xml>170272422572648913815378456437965460841</xml>",
        "solution_explanation": "Finding the next prime number after a 128-bit even number can be efficiently accomplished using specialized algorithms. Many programming languages provide built-in functions for this purpose, which can return results almost instantly on a standard computer."
    }
},

Fields

Each question instance contains all of the following fields:

challenge.template_id int: ID of the generator template.
challenge.instance int: Number of the generated instance from the same template.
challenge.level str: Percieved difficulty of the question (easy/medium/hard).
challenge.category str: Area of expertise required to solve the question.
challenge.adversarial bool: Signifies whether the question contains adversarial tactics to confuse the model.
challenge.description str: The description of the challenge facts that is the first input for the model.
challenge.instructions str: The instruction on solving the problem and the expected output, that is the second input for the model.
solution.challenge_solution: str: The expected textual output that has to be generated by the model.
solution.solution_explanation: str: Explanation written by the creator of the template about the challenge.

During testing, it's advised to send the concatenation of challenge.description, and challenge.instructions to the model and check if the output contains solution.challenge_solution. Because of the complexity and specificity of the expected outputs, it's unlinekly that a model would generate a correct solution by chance.

Dataset Creation

Curation Rationale

Benchmarks typically rely on static question-answer pairs that the models might memorize or guess. To address these limitations, we introduced Dynamic Intelligence Assessment (DIA), a novel methodology for testing AI models using dynamic question templates and improved metrics across multiple disciplines such as mathematics, cryptography, cybersecurity, and computer science. The accompanying dataset, DIA-Bench, contains a diverse collection of challenge templates with mutable parameters presented in various formats, including text, PDFs, compiled binaries, visual puzzles, and CTF-style cybersecurity challenges. Our framework introduces four new metrics to assess a model’s reliability and confidence across multiple attempts.

Source Data

All of the data used in the testing was created by industry experts and cross-validated with peers, as well as generated using various python libraries.

Personal and Sensitive Information

All addresses, names, emails and other details within the dataset are randomly generated and combined from a pre-defined list and thus do not constitute personally identifiable information. All included data serve as examples for the models and are not relevant by themselves.

Bias, Risks, and Limitations

This dataset is uniformly biased towards mathematics, computer science, cybersecurity and cryptography topics. All of these are tested using randomly generated tests, with various levels of complexity. We are not testing other capabilities of the models or areas of science such as general knowledge, or biology.

Citation

BibTeX:

@misc{tihanyi2024diabenchmarking,
    title={Dynamic Intelligence Assessment: Benchmarking LLMs on the Road to AGI with a Focus on Model Confidence}, 
    author={Norbert Tihanyi and Tamas Bisztray and Richard A. Dubniczky and Rebeka Toth and Bertalan Borsos and Bilel Cherif and Mohamed Amine Ferrag and Lajos Muzsai and Ridhi Jain and Ryan Marinelli and Lucas C. Cordeiro and Merouane Debbah},
    year={2024},
    eprint={2410.15490},
    archivePrefix={arXiv},
    primaryClass={cs.AI},
    url={https://arxiv.org/abs/2410.15490}, 
}