Skip to content
intermediateevaluation

MMLU-Pro - Tiered Multi-Subject Evaluation

Tiered evaluation for multi-subject question answering, based on MMLU-Pro (Wang et al., NeurIPS 2024). Annotators verify answers to challenging 10-option multiple choice questions across STEM and humanities subjects, using a tiered annotation scheme for topic and subtopic categorization.

About this dataset

MMLU-Pro is a multiple-choice benchmark that extends the original MMLU across academic and professional subjects. It was built by Yubo Wang and colleagues (University of Waterloo and collaborators) and presented as a spotlight at the NeurIPS 2024 Datasets and Benchmarks track. It exists because leading models had begun to saturate MMLU, leaving little room to distinguish strong systems.

The questions were drawn and revised from MMLU and other sources, with the answer set expanded from four options to ten and trivial or noisy items removed. Each question has a single correct answer among up to ten options (A through J). The annotation task is to categorize a question by subject and select the correct option.

The dataset holds 12,032 questions across 14 disciplines: math, physics, chemistry, law, engineering, economics, health, psychology, business, biology, philosophy, computer science, history, and an 'other' category. Adding options and pruning weak items lowered model accuracy by 16 to 33 percent relative to MMLU and made scores more stable across prompt wordings (about 2 percent variation versus 4 to 5 percent on MMLU).

The Potato config below reproduces this task with a tiered annotation scheme that captures a topic and a dependent subtopic, plus a radio scheme for the ten answer options (A-J). Use it to sort questions into a subject hierarchy while verifying the correct answer.

Released
NeurIPS 2024 D&B (spotlight)
Questions
12,032
Disciplines
14
Answer options
Up to 10 (A-J)
Accuracy drop vs MMLU
16-33%
Prompt sensitivity
~2% (vs 4-5% in MMLU)
Submit

Configuration Fileconfig.yaml

This Potato config reproduces the annotation task. Save it as config.yaml and run potato start config.yaml to try it.

yaml
# MMLU-Pro - Tiered Multi-Subject Evaluation
# Based on Wang et al., NeurIPS 2024
# Paper: https://arxiv.org/abs/2406.01574
# Dataset: https://huggingface.co/datasets/TIGER-Lab/MMLU-Pro
#
# MMLU-Pro extends MMLU with 10 answer options (A-J) instead of 4,
# making it significantly more challenging. This task uses a tiered
# annotation scheme to categorize questions by topic and subtopic,
# alongside answer selection.
#
# The tiered annotation allows organizing questions hierarchically:
# - Topic: The broad subject area (e.g., Biology, Physics, History)
# - Subtopic: A more specific area within the topic
#
# Answer options: A through J (10 choices per question)

annotation_task_name: "MMLU-Pro: Tiered Multi-Subject Evaluation"
task_dir: "."

data_files:
  - sample-data.json

item_properties:
  id_key: "id"
  text_key: "text"

output_annotation_dir: "annotation_output/"
output_annotation_format: "json"

port: 8000
server_name: localhost

annotation_schemes:
  - annotation_type: tiered_annotation
    name: subject_classification
    description: "Classify the question by topic and subtopic using the tiered hierarchy"
    source_field: "audio_url"
    media_type: "audio"
    tiers:
      - name: "topic"
        tier_type: "independent"
      - name: "subtopic"
        tier_type: "dependent"
        parent_tier: "topic"
        constraint_type: "symbolic_association"

  - annotation_type: radio
    name: correct_answer
    description: "Select the correct answer from the 10 options (A-J)"
    labels:
      - "A"
      - "B"
      - "C"
      - "D"
      - "E"
      - "F"
      - "G"
      - "H"
      - "I"
      - "J"
    keyboard_shortcuts:
      "A": "1"
      "B": "2"
      "C": "3"
      "D": "4"
      "E": "5"
      "F": "6"
      "G": "7"
      "H": "8"
      "I": "9"
      "J": "0"

annotation_instructions: |
  You will evaluate challenging multiple-choice questions from MMLU-Pro.
  1. Read the question and all 10 answer options carefully.
  2. Classify the question by topic and subtopic using the tiered scheme.
  3. Select the single correct answer (A through J).
  4. These questions are intentionally difficult and may require expert knowledge.

html_layout: |
  <div style="padding: 15px; max-width: 800px; margin: auto;">
    <div style="background: #f0f9ff; border: 1px solid #bae6fd; border-radius: 8px; padding: 16px; margin-bottom: 16px;">
      <span style="display: inline-block; background: #0369a1; color: white; padding: 2px 10px; border-radius: 12px; font-size: 13px; margin-bottom: 8px;">{{subject}}</span>
      <p style="font-size: 16px; font-weight: 600; line-height: 1.6; margin: 8px 0 0 0;">{{text}}</p>
    </div>
    <div style="background: #f8fafc; border: 1px solid #e2e8f0; border-radius: 8px; padding: 16px; margin-bottom: 16px;">
      <strong style="color: #475569;">Answer Options:</strong>
      <p style="font-size: 15px; line-height: 1.8; margin: 8px 0 0 0; white-space: pre-line;">{{options}}</p>
    </div>
  </div>

allow_all_users: true
instances_per_annotator: 50
annotation_per_instance: 3
allow_skip: true
skip_reason_required: false

Sample Datasample-data.json

json
[
  {
    "id": "mmlu_pro_001",
    "text": "Which of the following best describes the role of topoisomerase II in DNA replication?",
    "options": "A. It unwinds the double helix ahead of the replication fork\nB. It synthesizes RNA primers for Okazaki fragments\nC. It relieves positive supercoiling by making transient double-strand breaks\nD. It joins Okazaki fragments on the lagging strand\nE. It proofreads newly synthesized DNA\nF. It degrades RNA primers after replication\nG. It adds telomeric sequences to chromosome ends\nH. It methylates newly synthesized DNA strands\nI. It prevents re-replication by licensing origins\nJ. It stabilizes single-stranded DNA at the replication fork",
    "subject": "Biology",
    "audio_url": ""
  },
  {
    "id": "mmlu_pro_002",
    "text": "A projectile is launched at an angle of 60 degrees above the horizontal with an initial speed of 50 m/s. Ignoring air resistance, what is the maximum height reached by the projectile?",
    "options": "A. 45.9 m\nB. 55.7 m\nC. 63.8 m\nD. 76.5 m\nE. 85.3 m\nF. 95.7 m\nG. 102.4 m\nH. 110.2 m\nI. 127.6 m\nJ. 143.1 m",
    "subject": "Physics",
    "audio_url": ""
  }
]

// ... and 8 more items

Get This Design

View on GitHub

Clone or download from the repository

Quick start:

git clone https://github.com/davidjurgens/potato-showcase.git
cd potato-showcase/preference-learning/mmlu-pro-tiered-eval
potato start config.yaml

Dataset & paper

Wang et al., NeurIPS 2024 (Datasets & Benchmarks)

Citation (BibTeX)

bibtex
@inproceedings{wang2024mmlupro,
    title = "{MMLU}-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmark",
    author = "Wang, Yubo and Ma, Xueguang and Zhang, Ge and Ni, Yuansheng and Chandra, Abhranil and Guo, Shiguang and Ren, Weiming and Arulraj, Aaran and He, Xuan and Jiang, Ziyan and Li, Tianle and Ku, Max and Wang, Kai and Zhuang, Alex and Fan, Rongqi and Yue, Xiang and Chen, Wenhu",
    booktitle = "Advances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track",
    year = "2024",
    url = "https://arxiv.org/abs/2406.01574"
}

Details

Annotation Types

tiered_annotationradio

Domain

NLPEducation

Use Cases

Question AnsweringBenchmark EvaluationMulti-Subject Assessment

Tags

mmlu-promultiple-choicetieredevaluationstemhumanitiesneurips2024

Found an issue or want to improve this design?

Open an Issue