GAMUT
收藏资源简介:
GAMUT(Grounded Assessment of Multimodal Factuality)是一个多模态的日常深度研究基准,也是一个用于评估开放式、长文本生成的两级元量规框架。该数据集旨在解决基于量规的评价中的一个核心矛盾:评估开放式生成需要结构化的标准(因为完整答案通常不能分解为独立且必需的事实,例如答案可能来自开放的有效选项池,或者描述顺序重要的过程),但大语言模型评委在评分扁平、二元的检查项时更为可靠且方差更低。GAMUT通过两级量规解决了这个问题:首先,一个结构化的元量规用于制定评分标准(即完整答案所需的内容),然后该元量规被机械地编译成扁平化的二元检查项,供评委逐一评分。数据集包含1,813个基于真实可穿戴设备图像的问题,涵盖10个不同的领域。这些问题复杂且贴近现实世界,需要进行多步骤的信息收集和多段落的答案综合。问题和量规通过多轮人类与大语言模型协作的标注过程构建,每个量规都有引用的网络证据支持,并经过专家标注者的验证。该基准具有挑战性和区分性,评估的14个专有和开源模型中,最佳模型(Gemini 3.1 Pro)的得分仅为58.7%,表明基准远未饱和,且不同模型间得分差距显著,排名对不同评委模型的选择具有鲁棒性。数据集字段包括:`session_id`(唯一标识符,也是连接CRAG-MM数据集中图像的键)、`question`(关于图像的问题)和`rubrics`(包含答案关键/有价值/上下文信息及来源片段的两级元量规结构)。数据集提供两个测试分割:主多模态基准的`test`集(1,813个示例,每个问题基于一张图像)和用于评估无图像输入模型的纯文本变体`test_text_only`集(1,806个示例,字段相同)。请注意,图像本身不包含在此数据集中,需要通过`session_id`从CRAG-MM数据集中获取。该数据集采用CC-BY-NC-4.0许可证,主要用于基准测试。
GAMUT (Grounded Assessment of Multimodal Factuality) is a multimodal, daily-depth research benchmark and a two-level meta-rubric framework for evaluating open-ended, long-form generation. The dataset aims to address a core tension in rubric-based evaluation: assessing open-ended generation requires structured criteria (as complete answers often cannot be broken into independent, necessary facts, e.g., answers may come from an open pool of valid options or describe processes where order matters), but large language model judges are more reliable and lower-variance when scoring flat, binary checklist items. GAMUT resolves this via a two-level rubric: first, a structured meta-rubric for defining scoring criteria (i.e., what content is needed for a complete answer), which is then mechanically compiled into flattened binary checklist items for judges to score individually. The dataset contains 1,813 questions based on real wearable device images, covering 10 distinct domains. These questions are complex and grounded in the real world, requiring multi-step information gathering and multi-paragraph answer synthesis. Questions and rubrics are constructed through a multi-round human-LLM collaborative annotation process, with each rubric supported by cited web evidence and validated by expert annotators. The benchmark is challenging and discriminative, with the best-performing model (Gemini 3.1 Pro) scoring only 58.7% among 14 evaluated proprietary and open-source models, indicating that the benchmark is far from saturated, and score gaps between models are significant, with rankings robust to the choice of judge model. Dataset fields include: `session_id` (a unique identifier, also a key to link images from the CRAG-MM dataset), `question` (a question about the image), and `rubrics` (a two-level meta-rubric structure containing key/valuable/contextual information for the answer and source snippets). The dataset provides two test splits: the `test` set for the main multimodal benchmark (1,813 examples, each question based on one image) and a text-only variant `test_text_only` for evaluating models without image input (1,806 examples, same fields). Note that the images themselves are not included in this dataset and must be obtained via `session_id` from the CRAG-MM dataset. The dataset is licensed under CC-BY-NC-4.0 and is primarily intended for benchmarking.
数据集概述
GAMUT(Grounded Assessment of Multimodal Factuality) 是一个多模态日常深度研究基准,旨在评估开放式的长文本生成任务。该数据集由Meta(Facebook)发布,采用 CC-BY-NC 4.0 许可证。
核心特征
- 双重元评分框架:解决基于评分的评估中的结构性与评分可靠性之间的张力。
- 结构化的元评分:描述完整答案所需的标准。
- 扁平的二元检查:将元评分机械编译成一系列独立的二元检查项,使评估更稳定、方差更低。
- 日常深度研究基准:
- 1,813 个基于图像的问答示例,覆盖 10 个不同领域。
- 问题复杂,需多步信息收集和多段落答案合成。
- 经过多轮人工与LLM协同标注,每项评分均有引用的网络证据支持,并由专家验证。
任务与语言
- 任务类别:视觉问答(visual-question-answering)、问答(question-answering)
- 语言:英语(en)
数据集规模
- 规模:1K < n < 10K
数据集字段
| 字段 | 类型 | 描述 |
|---|---|---|
session_id |
字符串 | 唯一ID,也是关联CRAG-MM中图像的键 |
question |
字符串 | 关于图像的问题 |
rubrics |
结构体 | 双重元评分(包括“答案关键”、“有价值”、“上下文”),附带来源片段 |
数据分割
test(1,813 个示例):主要的多模态基准,每个问题与一张图像关联。test_text_only(1,806 个示例):纯文本变体,用于评估无图像输入的模型,字段相同。
图像来源
- 图像不包含在本数据集中,通过
session_id引用 CRAG-MM 数据集。 - 图像下载脚本及评估工具链位于配套的 GitHub 仓库:https://github.com/facebookresearch/GAMUT
许可证
- 数据集采用 CC-BY-NC 4.0 许可证,仅用于基准测试目的。
- 第三方内容受其自身许可证约束。
参考文献
bibtex @misc{chen2026gamut, title={Two-Level Meta-Rubrics for Evaluating Open-Ended Generation: {GAMUT}, a Benchmark for Factual Completeness}, author={Xilun Chen and Zhaleh Feizollahi and Ross Goodwin and Seungwhan Moon and Scott Yih and Pinar Donmez and Babak Damavandi and Luna Dong}, year={2026}, eprint={2607.19322}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2607.19322}, }




