hazyresearch/MMLU_with_Llama_3.1_70B_Instruct_v1

MMLU with Llama-3.1-70B-Instruct

This dataset contains multiple-choice questions from the MMLU benchmark with 100 candidate responses generated by Llama-3.1-70B-Instruct for each problem. Each response has been evaluated for correctness using a mixture of GPT-4o-mini and procedural Python code to robustly parse different answer formats, and scored by multiple reward models (scalar values) and LM judges (boolean verdicts).

Dataset Structure

• Split: Single split named "data"
• Num rows: 719 MMLU problems across diverse academic subjects
• Generations per query: 100

Key Fields

Field	Type	Description
instruction	str	Prompt given to Llama 3.1 70B Instruct
samples	List[str]	Model-generated answers (100 per problem)
extracted_answers	List[str]	Final answers extracted from completions (A, B, C, or D)
answer_correct	List[bool]	Whether each extracted answer matches the correct choice
*_verdicts	Dict[str, List[float]]	Binary signals from verifier models (e.g., LM judges)
*_scores	Dict[str, List[float]]	Scalar scores from reward models

Example Entry

{
  "instruction": "The following is a multiple choice question. Answer with the letter of the correct choice.\n\nQuestion: What is the capital of France?\nA. Berlin\nB. Madrid\nC. Paris\nD. Rome\n\nAnswer:",
  "samples": ["Looking at this question about European capitals...", "The answer is C", "I believe the answer is C", ...],
  "extracted_answers": ["C", "C", "B", ...],
  "answer_correct": [true, true, false, ...],
  "Llama-3.3-70B-Instruct_verdicts": [1.0, 1.0, 0.0, ...],
  "GRMGemma_scores": [0.78, 0.82, 0.34, ...],
  ...
}

Quick Start

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("hazyresearch/MMLU_with_Llama_3.1_70B_Instruct_v1")["data"]

# Get the first problem
problem = dataset[0]
print(f"Problem: {problem['instruction']}")

# Select the best response using pre-computed Weaver scores
best_idx = max(range(len(problem['weaver_scores'])), key=lambda i: problem['weaver_scores'][i])
best_response = problem['samples'][best_idx]
print(f"\nBest response (Weaver): {best_response}")

# Check if it's actually correct
print(f"Is correct: {problem['answer_correct'][best_idx]}")

Source

Original MMLU problems from cais/mmlu.

Usage with Weaver

This dataset can be used with the Weaver framework for training and evaluating verifier aggregation methods. See the repository for detailed instructions on reproducing paper results.

Citation

@misc{saadfalcon2025shrinkinggenerationverificationgapweak,
      title={Shrinking the Generation-Verification Gap with Weak Verifiers}, 
      author={Jon Saad-Falcon and E. Kelly Buchanan and Mayee F. Chen and Tzu-Heng Huang and Brendan McLaughlin and Tanvir Bhathal and Shang Zhu and Ben Athiwaratkun and Frederic Sala and Scott Linderman and Azalia Mirhoseini and Christopher Ré},
      year={2025},
      eprint={2506.18203},
      archivePrefix={arXiv},
      primaryClass={cs.CR},
      url={https://arxiv.org/abs/2506.18203}, 
}