human_study
收藏资源简介:
该数据集名为Human Study,是一个用于文本生成任务的小型英语数据集。其内容源自snap-stanford/user_study_annotations数据集,由Reddit上的自由形式人类回复构成,原始Reddit帖子文本来自HumanLM作者提供的reddit_post_dict_testset.json文件。每条数据样本包含两个核心字段:prompt字段存储原始的Reddit帖子文本,target字段存储由人类研究参与者针对该帖子撰写的自由形式回复。数据集明确排除了模型生成的回复、生成的人物设定、比较判断以及工作者标识符等信息。在数据处理上,基于特定的源注释修订版本和随机种子,对数据进行了去重处理,确保每个唯一的post_id仅随机保留一条注释,且post_id在整个数据集中具有唯一性。数据集被划分为两个部分:少样本训练集(train)和测试集(test)。训练集由5条最紧凑(旨在适应小模型上下文窗口)的选定数据行组成;测试集则包含其余所有唯一的帖子,共59条数据。数据集的提示格式经过特殊设计:以Post:开头展示原始帖子,后接Response:提示词(该词末尾精确包含一个ASCII空格,以便与lm-eval等评估工具兼容,实现在自然令牌边界开始生成),target即为紧接其后的参与者原始回复文本。
The dataset named Human Study is a small English dataset for text generation tasks. Its content is derived from the snap-stanford/user_study_annotations dataset, consisting of free-form human responses from Reddit. The original Reddit post text comes from the reddit_post_dict_testset.json file provided by the HumanLM authors. Each data sample includes two core fields: the prompt field stores the original Reddit post text, and the target field stores the free-form responses written by human study participants for that post. The dataset explicitly excludes model-generated responses, generated personas, comparative judgments, and worker identifiers. In data processing, deduplication is performed based on specific source annotation revision versions and random seeds, ensuring that only one annotation is randomly retained for each unique post_id, and post_id is unique throughout the dataset. The dataset is divided into two parts: a few-shot training set (train) and a test set (test). The training set consists of 5 selected data rows that are most compact (aimed at fitting small model context windows); the test set contains all other unique posts, totaling 59 data points. The prompt format of the dataset is specially designed: it starts with Post: to display the original post, followed by the Response: prompt (which ends with exactly one ASCII space to ensure compatibility with evaluation tools like lm-eval, enabling generation at natural token boundaries), and the target is the participants original response text immediately following.
数据集概述:human_study
- 许可证:Apache-2.0
- 任务类别:文本生成 (text-generation)
- 语言:英语 (en)
- 规模:n < 1K(总计少于1000条)
- 来源:
- 基于
snap-stanford/user_study_annotations数据集的自由形式人类Reddit回复。 - 原始Reddit帖子文本来源于HumanLM作者的
reddit_post_dict_testset.json。
- 基于
数据内容
- 每一行包含:
prompt:原始Reddit帖子的文本。target:由人类研究参与者撰写的回复。
- 排除了模型生成的回复、生成的人物画像、比较判断以及工作人员标识符。
去重与数据划分
- 源标注修订版本:
b19978763f7a1e4fcf71c2facfa65af0136cc425 - 随机种子:
42 - 去重规则:每个唯一的
post_id随机选择一条标注。 post_id在整个派生数据集中唯一。- 划分方式:
- 训练集 (train):选取最紧凑的5行作为少样本训练集,以适应当小模型上下文窗口。
- 测试集 (test):所有剩余的唯一条目构成评分测试集。
- 数据量:
- 训练集:5行
- 测试集:59行
提示词格式
-
固定提示模板:
Post: <原始Reddit帖子>
Response:<SPACE>
-
说明:
Response:提示以恰好一个ASCII空格结尾,以便lm-eval可以附加少样本目标,并在自然token边界开始生成。 -
目标:参与者的原始自由形式回复。




