generations-qwen3-8b-rmu-bm25-10b-rebuttal
收藏资源简介:
该数据集是一个多配置集合,主要用于评估语言模型在复杂推理任务上的性能。核心数据来源于两个知名基准:ARC(AI2推理挑战赛)和BBH(Big-Bench Hard)。数据集包含多个具体任务配置,例如ARC挑战赛配置,以及一系列带有“bbh_cot_fewshot_”前缀的BBH任务配置(如布尔表达式、因果判断、日期理解、形式谬误、逻辑推理、多步算术、对象计数等),这些配置均采用了思维链(Chain-of-Thought)和少样本(Few-shot)的提示设置。每个数据样本结构丰富,通常包含原始问题文档(如问题文本、选择题选项和答案键)、目标输出、用于生成模型响应的参数配置、模型生成的多轮原始响应、经过过滤的响应、用于唯一标识的哈希值以及一个评估分数。数据规模因配置而异,例如ARC挑战赛配置包含1172个训练样本,而大多数BBH配置包含250个训练样本。该数据集适用于研究语言模型的推理能力、评估模型在多样化任务上的泛化性能,以及对模型生成过程进行深入分析。
This dataset is a multi-configuration collection primarily used to evaluate the performance of language models on complex reasoning tasks. The core data is sourced from two well-known benchmarks: ARC (AI2 Reasoning Challenge) and BBH (Big-Bench Hard). The dataset includes multiple specific task configurations, such as the ARC challenge configuration, and a series of BBH task configurations prefixed with bbh_cot_fewshot_ (e.g., boolean expressions, causal judgment, date understanding, formal fallacies, logical reasoning, multi-step arithmetic, object counting, etc.), all of which employ Chain-of-Thought and Few-shot prompting settings. Each data sample has a rich structure, typically containing the original question document (e.g., question text, multiple-choice options, and answer key), target output, parameter configurations for generating model responses, raw multi-turn responses generated by models, filtered responses, a hash for unique identification, and an evaluation score. The data scale varies by configuration; for example, the ARC challenge configuration includes 1172 training samples, while most BBH configurations include 250 training samples. This dataset is suitable for researching the reasoning capabilities of language models, evaluating their generalization performance across diverse tasks, and conducting in-depth analysis of the model generation process.
数据集概述
该数据集为 anime-sh/generations-qwen3-8b-rmu-bm25-10b-rebuttal,是 Hugging Face 上的一个数据集,包含多个配置(config),每个配置对应不同的推理任务评测数据。
数据集内容
数据集主要由多个配置组成,每个配置对应一个具体的推理任务,具体包括:
- arc_challenge(AI2 推理挑战)
- bbh_cot_fewshot 系列,涵盖多个推理任务,例如:
- boolean_expressions(布尔表达式)
- causal_judgement(因果判断)
- date_understanding(日期理解)
- disambiguation_qa(消歧问答)
- dyck_languages(Dyck 语言)
- formal_fallacies(形式谬误)
- geometric_shapes(几何形状)
- hyperbaton(修辞倒装)
- logical_deduction_five_objects(五对象逻辑推理)
- logical_deduction_seven_objects(七对象逻辑推理)
- logical_deduction_three_objects(三对象逻辑推理)
- movie_recommendation(电影推荐)
- multistep_arithmetic_two(多步算术)
- navigate(导航)
- object_counting(物体计数)
- penguins_in_a_table(表格中的企鹅)
- reasoning_about_colored_objects(彩色物体推理)
- ruin_names(损毁名称)
数据结构
各配置的数据特征如下:
-
arc_challenge:
doc_id(int64):文档 IDdoc:包含answerKey、choices(label 和 text)、id、question等原始评测信息target(string):目标输出arguments:包含 5 组生成参数(gen_args_0 至 gen_args_4),每组含两个参数resps:嵌套列表结构,存模型原始响应filtered_resps:过滤后的响应filter、metrics、doc_hash、prompt_hash、target_hash、score等元数据
-
bbh_cot_fewshot 系列:
doc_id(int64)doc:包含input和target字段target(string):目标输出arguments:生成参数,包含arg_0(string)和arg_1(内含do_sample、max_gen_toks、temperature、until等采样配置)resps:模型响应列表filtered_resps:过滤后的响应列表filter、metrics、doc_hash、prompt_hash、target_hash、bypass、score等元数据
各配置数据规模
所有配置均仅包含 train 分割。以下为各配置的样本数量与大小示例:
| 配置名称 | 样本数 | 数据集大小(字节) | 下载大小(字节) |
|---|---|---|---|
| arc_challenge | 1,172 | 1,903,272 | 1,729,033 |
| bbh_cot_fewshot_boolean_expressions | 250 | 696,817 | 686,773 |
| bbh_cot_fewshot_causal_judgement | 187 | 1,248,292 | 1,240,870 |
| bbh_cot_fewshot_date_understanding | 250 | 624,494 | 612,634 |
| bbh_cot_fewshot_disambiguation_qa | 250 | 1,397,907 | 1,398,329 |
| bbh_cot_fewshot_dyck_languages | 250 | 999,319 | 997,882 |
| bbh_cot_fewshot_formal_fallacies | 250 | 1,766,553 | 1,751,353 |
| bbh_cot_fewshot_geometric_shapes | 250 | 1,842,570 | 1,833,163 |
| bbh_cot_fewshot_hyperbaton | 250 | 1,183,277 | 1,180,939 |
| bbh_cot_fewshot_logical_deduction_five_objects | 250 | 1,343,420 | 1,341,006 |
| bbh_cot_fewshot_logical_deduction_seven_objects | 250 | 1,533,663 | 1,543,001 |
| bbh_cot_fewshot_logical_deduction_three_objects | 250 | 1,098,087 | 1,095,022 |
| bbh_cot_fewshot_movie_recommendation | 250 | 892,361 | 885,178 |
| bbh_cot_fewshot_multistep_arithmetic_two | 250 | 876,072 | 881,541 |
| bbh_cot_fewshot_navigate | 250 | 827,206 | 819,266 |
| bbh_cot_fewshot_object_counting | 250 | 565,339 | 554,402 |
| bbh_cot_fewshot_penguins_in_a_table | 146 | 607,633 | 615,440 |
| bbh_cot_fewshot_reasoning_about_colored_objects | 250 | 966,205 | 960,272 |
| bbh_cot_fewshot_ruin_names(信息不完整) | - | - | - |
总结
该数据集是用于评估 Qwen3-8B 模型在多种推理任务上的生成结果数据集,覆盖 ARC-Challenge 以及 BIG-Bench Hard(BBH)中的多项推理任务(如逻辑推理、日期理解、数学计算、导航等)。每个配置中包含模型的原始响应、过滤后的响应以及相应的评分与哈希信息,适用于分析模型在不同推理场景下的表现与评估结果。




