math500-rollouts
收藏资源简介:
该数据集是一个用于评估语言模型性能的结构化数据集。数据集包含20个训练样本,每个样本代表一个具体的任务实例。数据集中每个样本包含以下核心信息:任务标识符(id)、难度级别(level)、任务分组(group)、输入提示词(prompt)以及标准答案(ground_truth)。此外,数据集记录了模型对该提示的多个响应(responses),包括每个响应的文本内容、token数量(response_tokens)和字符数量(response_chars),以及每个响应是否正确(is_correct)的评估结果。数据集还提供了聚合的性能指标:平均准确率(mean_accuracy)、是否至少有一个正确响应(pass_at_least_one)、平均响应token数(mean_tokens)和平均响应字符数(mean_chars)。特别地,数据集包含了详细的推理效率指标:预填充时间(prefill_times)和解码时间(decode_times)列表,对应的token/秒(prefill_tok_s_list, decode_tok_s_list)和字符/秒(prefill_char_s_list, decode_char_s_list)速率列表,以及这些速率指标的平均值(mean_prefill_tok_s, mean_decode_tok_s, mean_prefill_char_s, mean_decode_char_s)。该数据集适用于评估语言模型在特定任务(如代码生成、问题解答等)上的准确性、响应质量和推理效率,为模型性能的多维度分析提供数据支持。
This is a structured dataset intended for evaluating the performance of language models. The dataset consists of 20 training samples, where each sample represents a concrete task instance. Each core sample contains the following key information: task identifier (id), difficulty level (level), task group (group), input prompt (prompt), and standard answer (ground_truth). In addition, the dataset records multiple model responses to the given prompt, including the text content of each response, the token count (response_tokens) and character count (response_chars) of each response, as well as the evaluation result indicating whether each response is correct (is_correct). The dataset also provides aggregated performance metrics: mean accuracy (mean_accuracy), whether at least one correct response exists (pass_at_least_one), average number of response tokens (mean_tokens), and average number of response characters (mean_chars). Notably, the dataset includes detailed inference efficiency metrics: lists of prefill times and decode times, corresponding token-per-second lists (prefill_tok_s_list, decode_tok_s_list) and character-per-second lists (prefill_char_s_list, decode_char_s_list), as well as the average values of these rate metrics (mean_prefill_tok_s, mean_decode_tok_s, mean_prefill_char_s, mean_decode_char_s). This dataset can be used to evaluate the accuracy, response quality, and inference efficiency of language models across specific tasks such as code generation and question answering, providing data support for multi-dimensional analysis of model performance.
该数据集来自Hugging Face上的deepnevro/math500-rollouts,是一个面向数学推理任务的数据集。数据集仅包含一个训练集(train),样本数量为1条,整体数据集大小为27,371字节,下载大小为38,303字节。
数据集包含以下字段:
id:整数类型,样本标识符。level:字符串类型,数学问题难度等级。group:字符串类型,样本所属分组。prompt:字符串类型,数学问题提示内容。ground_truth:字符串类型,正确答案。responses:字符串列表,模型生成的多个回答。response_tokens:整数列表,每个回答对应的token数量。response_chars:整数列表,每个回答对应的字符数量。is_correct:布尔列表,每个回答是否正确。mean_accuracy:整数类型,平均准确率。pass_at_least_one:布尔类型,是否至少有一个正确回答。mean_tokens:整数类型,所有回答的平均token数。mean_chars:整数类型,所有回答的平均字符数。prefill_times:浮点数列表,每次生成的预填充阶段耗时。decode_times:浮点数列表,每次生成的解码阶段耗时。prefill_tok_s_list:浮点数列表,预填充阶段每秒处理的token数。decode_tok_s_list:浮点数列表,解码阶段每秒生成的token数。mean_prefill_tok_s:浮点数类型,预填充阶段平均每秒处理token数。mean_decode_tok_s:浮点数类型,解码阶段平均每秒生成token数。mean_prefill_char_s:浮点数类型,预填充阶段平均每秒处理字符数。mean_decode_char_s:浮点数类型,解码阶段平均每秒生成字符数。
该数据集适用于评估数学推理模型的生成性能、准确率以及生成效率(包括速度与资源消耗)。




