cot-qwen3-8b-multiscale-inference
收藏资源简介:
该数据集是一个大规模、结构化的文本数据集,专门为语言模型的训练与评估而构建,核心聚焦于问答、推理与上下文理解任务。数据集包含 88,843 个训练样本。每个样本由丰富的字段构成,主要包括:原始文档或上下文文本(doc, context)、问题(question)、正确答案(correct_answer, target_response)、以及用于引导模型输出的提示词(verbalizer_prompt)。此外,数据集还包含了详细的元数据和过程信息,例如:问题的领域(domain)、学科主题(subject)、数据来源(source, source_tag)、生成模型(generator_model)、以及问题所属的类别(compqa_category, scope)。为了支持复杂的推理分析,数据提供了多种理性分析字段,如文档选择依据(doc_choice_rationale)和划分依据(split_rationale),并标注了文本在不同粒度(句子、段落、全文)上的划分信息(row_splits)。数据质量经过多维度检查,包括协同性检查(synergy_check)、分布检查(distribution_check)、难度检查(difficulty_check)和计算层面评估(computational_aspect)。该数据集适用于多项自然语言处理任务,如开放域问答、阅读 comprehension、推理能力评测、提示工程优化以及语言模型在科学、学术等多领域知识上的能力评估。
This dataset is a large-scale, structured text dataset specifically constructed for the training and evaluation of language models, with a core focus on question answering, reasoning, and context understanding tasks. The dataset contains 88,843 training samples. Each sample includes a rich set of comprehensive fields, primarily containing: original document or context text (doc, context), question, correct answer (correct_answer, target_response), and verbalizer_prompt used to guide model outputs. In addition, the dataset provides detailed metadata and procedural information, including: the domain of the question, subject, data source (source, source_tag), generator_model, and the category of the question (compqa_category, scope). To enable complex reasoning analysis, the dataset offers multiple rational analysis fields, such as doc_choice_rationale and split_rationale, and annotates text partitioning information at different granularities (sentence, paragraph, full text) via the field row_splits. The data quality of this dataset has been verified through multi-dimensional checks, including synergy_check, distribution_check, difficulty_check, and computational_aspect evaluation. This dataset is suitable for a wide range of natural language processing tasks, including open-domain question answering, reading comprehension, reasoning capability evaluation, prompt engineering optimization, and evaluation of language models' capabilities on multi-domain knowledge such as scientific and academic fields.
- 数据集名称:cot-qwen3-8b-multiscale-inference
- 数据集页面:https://huggingface.co/datasets/cds-jb/cot-qwen3-8b-multiscale-inference
- 数据集大小:约 1.05 GB(下载大小约 182 MB)
- 数据规模:训练集共 88,843 条样本
- 数据格式:仅包含训练集(train),无验证集或测试集
- 数据结构:每条样本包含 33 个字段,涵盖文档、提示词、上下文、原子文本、目标响应、正确/错误答案、领域、主题、问题、推理依据、分割依据、评分(句级、段级、整体)、难度检查、计算方面、来源标签等
- 主要用途:用于多尺度推理(multiscale inference)相关的问答、文档理解或链式推理(chain-of-thought)任务
- 字段说明:
- 文档相关:doc、doc_id、doc_idx、doc_choice_rationale
- 提示与上下文:verbalizer_prompt、context、atom_text、question
- 答案与目标:target_response、correct_answer、incorrect_plausible_answer、target、typicality
- 分割与评分:split_rationale、row_splits(含 sentence、paragraph、whole 分数)、split_char_offset
- 语义标签:domain、subject、compqa_category、scope、source、source_tag
- 生成与检查:generator_model、synergy_check、distribution_check、difficulty_check、computational_aspect
- 其他索引:row_seed、slot_idx、abs_index、signed_index、question_track、ipa_dropped_reason




