augmented-sample-math-full
收藏资源简介:
该数据集用于评估模型在问题排列(permutation)下的性能变化。它包含多个字段,记录了模型ID、数据集ID、问题ID、原始问题文本、排列类型、排列来源、模型在原始问题上的准确率(base_accuracy)、在排列后问题上的准确率(permuted_accuracy)、绝对准确率衰减(absolute_accuracy_decay)、相对准确率衰减(relative_accuracy_decay)、原始预测数量、排列后预测数量、导致性能衰减的排列数量以及具体导致衰减的排列列表。数据集包含702个验证样本,适用于分析模型对问题表述变化的鲁棒性,以及识别容易导致模型性能下降的特定问题变换模式。
This dataset is used to evaluate the performance changes of models under question permutations. It includes multiple fields, recording model ID, dataset ID, question ID, original question text, permutation type, permutation source, model accuracy on the original question (base_accuracy), accuracy on the permuted question (permuted_accuracy), absolute accuracy decay, relative accuracy decay, original prediction count, permuted prediction count, number of permutations causing performance decay, and a list of specific permutations that lead to decay. The dataset contains 702 validation samples and is suitable for analyzing the robustness of models to changes in question formulation and identifying specific transformation patterns that easily cause model performance degradation.
数据集概述
数据集名称:augmented-sample-math-full
数据集地址:https://huggingface.co/datasets/aimo-interp/augmented-sample-math-full
数据集结构
该数据集包含以下字段:
- model_id:模型标识符(字符串类型)
- dataset_id:数据集标识符(字符串类型)
- problem_id:问题标识符(字符串类型)
- original_problem:原始问题内容(字符串类型)
- permutation_type:排列类型(字符串类型)
- permutation_source:排列来源(字符串类型)
- base_accuracy:基准准确率(浮点数)
- permuted_accuracy:排列后准确率(浮点数)
- absolute_accuracy_decay:绝对准确率衰减(浮点数)
- relative_accuracy_decay:相对准确率衰减(浮点数)
- n_base_predictions:基准预测数量(整数)
- n_permuted_predictions:排列后预测数量(整数)
- n_detrimental_permutations:有害排列数量(整数)
- permutations_causing_decay:导致衰减的排列(字符串类型)
数据集划分
- 验证集:包含 702 个样本,数据大小为 341,081 字节
数据集配置
- 配置名称:default
- 数据文件:验证集数据位于
data/validation-*路径下
数据集大小
- 下载大小:71,268 字节
- 数据集总大小:341,081 字节




