GPQA-Diamond#

Overview#

GPQA (Graduate-Level Google-Proof Q&A) Diamond is a challenging benchmark of 198 multiple-choice questions written by domain experts in biology, physics, and chemistry. The questions are designed to be extremely difficult, requiring PhD-level expertise to answer correctly.

Task Description#

  • Task Type: Expert-Level Multiple-Choice Q&A

  • Input: Graduate-level science question with 4 choices

  • Output: Single correct answer letter (A, B, C, or D)

  • Domains: Biology, Physics, Chemistry

Key Features#

  • 198 questions written and validated by domain PhD experts

  • Questions are “Google-proof” - cannot be easily looked up

  • Designed to test deep domain knowledge and reasoning

  • Diamond subset represents the highest quality questions

  • Average human expert accuracy ~65%, non-expert ~34%

Evaluation Notes#

  • Default configuration uses 0-shot or 5-shot evaluation

  • Supports Chain-of-Thought (CoT) prompting for improved reasoning

  • Answer choices are randomly shuffled during evaluation

  • Only uses train split (validation set is private)

  • Challenging benchmark for measuring expert-level reasoning

Properties#

Property

Value

Benchmark Name

gpqa_diamond

Dataset ID

AI-ModelScope/gpqa_diamond

Paper

N/A

Tags

Knowledge, MCQ

Metrics

acc

Default Shots

0-shot

Evaluation Split

train

Data Statistics#

Metric

Value

Total Samples

198

Prompt Length (Mean)

841.15 chars

Prompt Length (Min/Max)

340 / 5845 chars

Sample Example#

Subset: default

{
  "input": [
    {
      "id": "82b448a9",
      "content": "Answer the following multiple choice question. The last line of your response should be of the following format: 'ANSWER: [LETTER]' (without quotes) where [LETTER] is one of A,B,C,D. Think step by step before answering.\n\nTwo quantum states wi ... [TRUNCATED] ...  and 10^-8 sec, respectively. We want to clearly distinguish these two energy levels. Which one of the following options could be their energy difference so that they can be clearly resolved?\n\n\nA) 10^-4 eV\nB) 10^-9 eV\nC) 10^-8 eV\nD) 10^-11 eV"
    }
  ],
  "choices": [
    "10^-4 eV",
    "10^-9 eV",
    "10^-8 eV",
    "10^-11 eV"
  ],
  "target": "A",
  "id": 0,
  "group_id": 0,
  "subset_key": "",
  "metadata": {
    "correct_answer": "10^-4 eV",
    "incorrect_answers": [
      "10^-11 eV",
      "10^-8 eV\n",
      "10^-9 eV"
    ]
  }
}

Note: Some content was truncated for display.

Prompt Template#

Prompt Template:

Answer the following multiple choice question. The last line of your response should be of the following format: 'ANSWER: [LETTER]' (without quotes) where [LETTER] is one of {letters}. Think step by step before answering.

{question}

{choices}

Usage#

Using CLI#

evalscope eval \
    --model YOUR_MODEL \
    --api-url OPENAI_API_COMPAT_URL \
    --api-key EMPTY_TOKEN \
    --datasets gpqa_diamond \
    --limit 10  # Remove this line for formal evaluation

Using Python#

from evalscope import run_task
from evalscope.config import TaskConfig

task_cfg = TaskConfig(
    model='YOUR_MODEL',
    api_url='OPENAI_API_COMPAT_URL',
    api_key='EMPTY_TOKEN',
    datasets=['gpqa_diamond'],
    limit=10,  # Remove this line for formal evaluation
)

run_task(task_cfg=task_cfg)