README.md

metadata

dataset_name: openmed-community/mmlu-5-options-rl-ready
tags:
  - MMLU
  - evaluation
  - DPO
  - RL
  - SFT
pretty_name: MMLU – 5-Options RL-Ready
license: mit
language:
  - en
task_categories:
  - multiple-choice
  - question-answering
  - reinforcement-learning
dataset_info:
  features:
    - name: question
      dtype: string
    - name: subject
      dtype: string
    - name: choices
      list: string
    - name: answer
      dtype: int64
    - name: task
      dtype: string
    - name: output
      dtype: string
    - name: options
      dtype: string
    - name: letter
      dtype: string
    - name: incorrect_letters
      list: string
    - name: incorrect_answers
      list: string
    - name: single_incorrect_answer
      dtype: string
    - name: system_prompt
      dtype: string
    - name: input
      dtype: string
  splits:
    - name: train
      num_bytes: 369405907
      num_examples: 97842
    - name: test
      num_bytes: 7551070
      num_examples: 2000
  download_size: 229394171
  dataset_size: 376956977
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*

MMLU – 5-Options RL-Ready

A standardized, RL-friendly remix of MMLU with explicit negatives and a unified five-option presentation string for each question. Ideal for DPO and other RL setups while remaining drop-in for classic multiple-choice evaluation.

What’s inside

• Splits & size: ~97.8k train + 2k test ≈ 99.8k total.

• Schema (core fields):

  • question: str
  • choices: list[str] (canonical options, typically 4 as in original MMLU)
  • answer: int (0-based index)
  • task: str (subject/task label; ~55 values)
  • output: str (correct option text)
  • options: str (single markdown-style block with (1)…(5) enumerated choices for unified 5-option prompts)
  • letter: str (correct letter tag)
  • incorrect_letters: list[str]
  • incorrect_answers: list[str]
  • single_incorrect_answer: str (one negative for pairwise prefs)
  • system_prompt: str (single default value)
  • input: str (ready-to-use user message text)

Note: The dataset provides both the original structured choices array (as in MMLU) and a five-option options string for standardized, list-variant prompting in RL pipelines.

Why it’s RL-ready

• Explicit negatives: incorrect_answers + single_incorrect_answer enable DPO, pairwise prefs, and contrastive training without extra preprocessing.
• Unified prompts: system_prompt + input and the five-option options string make it simple to build consistent chat-style prompts across frameworks.

Example record

{
  "question": "Which statement best describes the critics' reaction to the Segway?",
  "choices": ["Nothing but an electrical device.", "A disappointing engineering mistake.", "An expensive and disappointing invention.", "Disappointing, but still a successful device."],
  "answer": 3,
  "task": "miscellaneous",
  "output": "Disappointing, but still a successful device.",
  "options": "(1) ... (2) ... (3) ... (4) ... (5) ...",
  "letter": "(3)",
  "incorrect_letters": ["(1)", "(2)", "(4)", "(5)"],
  "incorrect_answers": ["...", "...", "...", "..."],
  "single_incorrect_answer": "...",
  "system_prompt": "You are a helpful tutor.",
  "input": "Choose the correct answer from the options below.\n\n<question + (1)…(5) options>"
}

Intended uses

• Evaluation of general reasoning on MMLU tasks with standardized five-option prompts.
• SFT with chat-style formatting.
• DPO / RL using explicit positive vs. negative pairs from single_incorrect_answer or full incorrect_answers.

Source & attribution

Derived from the original MMLU dataset by Hendrycks et al. (CAIS) cais/mmlu. Please cite the original work when using this derivative.