val-sample
收藏资源简介:
该数据集是一个用于评估语言模型在问题表述变化(排列)下性能鲁棒性的基准测试集合。数据集记录了多个模型在不同问题排列场景下的性能指标,包含模型标识、数据集标识、问题标识、原始问题文本、排列类型、排列来源等元信息。核心指标包括模型在原始问题上的基础准确率、在排列后问题上的准确率、绝对与相对的准确率衰减值,以及导致性能下降的排列数量统计。数据集还包含一个布尔字段指示模型是否在该排列测试中表现鲁棒。该数据集适用于研究模型对问题表述变化的敏感性、评估模型鲁棒性、以及分析不同排列策略对模型性能的影响。数据集包含一个验证集,共28个样本。
This dataset is a benchmark collection for evaluating the robustness of language models against variations and permutations of question formulations. It records performance metrics of multiple models across different question permutation scenarios, including meta-information such as model ID, dataset ID, question ID, original question text, permutation type, and permutation source. Core metrics include the baseline accuracy of the model on the original question, the accuracy on permuted questions, absolute and relative accuracy decay values, as well as statistics on the number of permutations that lead to performance degradation. The dataset also contains a boolean field indicating whether the model performed robustly in this permutation test. This dataset is suitable for researching the sensitivity of models to question formulation variations, evaluating model robustness, and analyzing the impact of different permutation strategies on model performance. The dataset includes a validation set consisting of 28 samples.
根据您提供的数据集详情页面信息,以下是该数据集的总结:
数据集概述
基本信息
- 数据集名称:val-sample
- 数据集地址:https://huggingface.co/datasets/aimo-interp/val-sample
- 数据集大小:下载大小为 23,788 字节,数据集总大小为 56,483 字节
- 数据划分:仅包含验证集(validation),共 28 个样本
数据特征
该数据集包含以下 15 个字段:
| 字段名称 | 数据类型 | 描述 |
|---|---|---|
| model_id | large_string | 模型标识 |
| dataset_id | large_string | 数据集标识 |
| problem_id | large_string | 问题标识 |
| original_problem | large_string | 原始问题 |
| permutation_type | large_string | 置换类型 |
| permutation_source | large_string | 置换来源 |
| base_accuracy | float64 | 基础准确率 |
| permuted_accuracy | float64 | 置换后准确率 |
| absolute_accuracy_decay | float64 | 绝对准确率衰减 |
| relative_accuracy_decay | float64 | 相对准确率衰减 |
| n_base_predictions | int64 | 基础预测数量 |
| n_permuted_predictions | int64 | 置换后预测数量 |
| n_detrimental_permutations | int64 | 有害置换数量 |
| permutations_causing_decay | large_string | 导致衰减的置换 |
| model_is_robust | bool | 模型是否鲁棒 |
数据集用途
该数据集主要用于评估模型在问题置换条件下的鲁棒性,通过对比基础准确率和置换后准确率,计算准确率衰减程度,从而判断模型是否对特定类型的置换具有鲁棒性。




