hazyresearch/MATH500_with_Llama_3.1_70B_Instruct_v1

本数据集名为MATH-500 with Llama-3.1-70B-Instruct，包含了来自MATH基准的500个数学推理问题，每个问题由Llama-3.1-70B-Instruct模型生成了100个候选答案。这些答案已经经过正确性评估，并由多个奖励模型和LM评判器进行了打分。数据集的结构包括多个字段，如instruction（指示）、samples（样本答案）、extracted_answers（提取的答案）、answer_correct（答案是否正确），以及来自不同模型的多个裁决和分数。该数据集遵循MIT许可证，并且可以配合Weaver框架用于训练和评估验证器聚合方法。

The MATH-500 with Llama-3.1-70B-Instruct dataset contains 500 mathematical reasoning problems from the MATH benchmark, each accompanied by 100 candidate responses generated by the Llama-3.1-70B-Instruct model. These responses have been evaluated for correctness and scored by various reward models and LM judges. The dataset is structured with fields such as instruction, samples, extracted_answers, answer_correct, along with multiple verdicts and scores from different models. The dataset is licensed under MIT and is compatible with the Weaver framework for training and evaluating verifier aggregation methods.

搜集汇总

背景与挑战

背景概述

该数据集包含500个来自MATH基准的数学推理问题，每个问题配有100个由Llama-3.1-70B-Instruct生成的候选回答，并经过多重验证机制（包括GPT-4o-mini和Python代码解析）评估正确性。数据集主要用于训练和评估验证器聚合方法，特别是与Weaver框架结合，以缩小生成与验证之间的差距。