FOXGLOVE
收藏资源简介:
FOXGLOVE是由伊利诺伊大学厄巴纳-香槟分校研究团队构建的写作反馈数据集,旨在系统比较专家与大型语言模型在议论文反馈上的差异。该数据集包含2,340条评论,基于69篇高中12年级议论文,涵盖696条训练有素的写作导师评论和1,644条来自GPT-5.2、Claude Sonnet 4.5等四个前沿模型的生成评论,每篇文章平均约423词。数据创建过程遵循统一的反馈协议,包括句子级锚定、目标导向标签和紧迫性排名标注,通过独立专家注释与模型提示生成并行收集。该数据集主要应用于教育技术领域,旨在探究人类与AI反馈在目标定位、文本锚定和优先级排序等维度的对齐与分歧,为智能化写作辅助工具的开发提供实证基础。
FOXGLOVE is a writing feedback dataset developed by the research team at the University of Illinois Urbana-Champaign, which aims to systematically compare the differences between human experts and large language models (LLMs) when providing argumentative writing feedback. The dataset consists of 2,340 comments derived from 69 12th-grade high school argumentative essays, including 696 comments from professionally trained writing tutors and 1,644 generated comments from four cutting-edge models such as GPT-5.2 and Claude Sonnet 4.5. Each essay has an average length of approximately 423 words. The dataset was constructed following a unified feedback protocol that covers sentence-level anchoring, goal-oriented labeling, and urgency ranking annotation, and was collected in parallel through independent expert annotations and model prompt-based generation. This dataset is primarily applied in the field of educational technology, with the goal of investigating the alignment and divergence between human and AI feedback across dimensions including goal orientation, text anchoring, and priority ranking, thereby providing an empirical foundation for the development of intelligent writing assistance tools.
数据集概述
FOXGLOVE(FOrmative feedback from eXperts, with Goal-oriented, Localized annotatiOns and eValuations of student Essays)是一个面向12年级议论文写作反馈的平行数据集,旨在系统比较人类专家与前沿大语言模型在写作反馈上的差异。
核心特点
- 数据来源:69篇来自PERSUADE 2.0语料库的12年级议论文
- 反馈来源:14名经过培训的写作指导教师,以及4个大语言模型(GPT-5.2、Claude Sonnet 4.5、Llama 3.3 70B、Qwen3-Next 80B)
- 数据规模:
- 人类反馈评论:696条
- 大语言模型反馈评论:1,644条
- 反馈评论总计:2,340条
- 专家质量评级:1,430条(覆盖7个维度,由2名额外评分员对105条评论进行评分)
反馈结构设计
每位反馈提供者(人类或模型)均按照统一纲要为每篇论文生成反馈:
- 句子级评论:锚定到连续句子,并标注为五种议论文目标之一(立场、论点、证据、反驳、反驳论证)
- 优先级排序:每篇论文中选出最重要的三条句子级评论(排名1–3)
- 全局评论:每篇论文一条自由文本评论,综合说明论文与议论文目标的契合度
文件结构
| 文件名 | 描述 |
|---|---|
essay_goals.csv |
包含12年级源论文全文及每位反馈提供者被要求关注的三个议论文目标 |
feedback.csv |
所有人类与模型的反馈评论,一行一条评论 |
ratings_global.csv |
专家对全局评论在6个质量维度的评分 |
ratings_items.csv |
专家对句子级评论在7个质量维度的评分 |
essay_goals.csv 字段
essay_id:匿名化论文标识符(关联至PERSUADE 2.0)Essay Contents:论文全文goal_1/goal_2/goal_3:呈现给反馈提供者的议论文目标
feedback.csv 字段
essay_id:评论对应的论文model:反馈来源(human或模型名称)reviewer_id:匿名化评论者标识符(如R1、R2)或模型名称goal_addressed:目标(立场、论点、证据、反驳、反驳论证或全局)urgency_rank:1(最紧急)、2、3(未排名则为空)feedback_content:自由文本反馈评论sentences_or_sections:评论锚定的论文文本高亮片段(全局评论为空)
ratings_global.csv 和 ratings_items.csv 字段
rater_id:匿名化评分员(如rater1、rater2)essay_id/real_model/reviewer_id:被评分反馈的身份信息anonymous_label:评分员看到的盲化标签(如Feedback Set C)feedback_id(仅句子级):对应feedback.csv中的行- 评分维度(1–5李克特量表):
accuracy_content、actionability、clarity、relevance、specificity、tone accuracy_position(仅句子级):锚定文本片段的位置准确性(不适用于全局评论)
许可协议
- 源论文来自PERSUADE 2.0语料库
- 已移除所有个人身份信息,教师与评分员仅以匿名标识符指代
- 数据采用CC BY-NC-SA 4.0 DEED Attribution-NonCommercial-ShareAlike 4.0 International许可(https://creativecommons.org/licenses/by-nc-sa/4.0/deed.en)




