MOSAIC
收藏资源简介:
MOSAIC是由Capital One研究团队开发的模块化评估基准,旨在对大型语言模型的指令遵循能力进行细粒度分析。该数据集包含4000条动态生成的文本提示,每条提示整合了内容生成任务、产品/服务描述及1-20个应用导向约束,覆盖格式、词汇、句法和语义四大类共21种约束类型。数据通过分层抽样构建,确保约束类型、列表长度和位置分布的平衡性。其创新性在于采用模块化设计将约束与核心任务解耦,支持对约束交互效应和位置偏置的独立研究,主要应用于评估LLM在复杂指令场景下的合规性表现,为提升模型可靠性提供诊断依据。
MOSAIC is a modular evaluation benchmark developed by the Capital One Research team, designed to conduct fine-grained analyses of the instruction-following capabilities of large language models (LLMs). This dataset contains 4,000 dynamically generated textual prompts, each integrating a content generation task, product/service descriptions, and 1 to 20 application-oriented constraints, covering a total of 21 constraint types across four major categories: format, lexicon, syntax, and semantics. The dataset is constructed via stratified sampling to ensure balanced distributions of constraint types, constraint list lengths, and constraint positions. Its core innovation lies in the modular design that decouples constraints from core tasks, enabling independent research on the interactive effects of constraints and positional biases. It is primarily applied to evaluate the compliance performance of LLMs in complex instruction scenarios, providing diagnostic evidence for improving model reliability.
数据集概述
数据集名称
MOSAIC (MOdular Synthetic Assessment of Instruction Compliance)
数据集简介
MOSAIC 是一个模块化框架,旨在对大型语言模型的指令遵循能力进行细粒度、独立的分析。该框架使用动态生成的数据集,其中包含多达20个面向应用的生成约束。
数据集生成方式
- 通过运行
python generate_modular_dataset.py脚本,可在项目目录下生成包含数据集的CSV文件。
相关环境配置
- Python版本:3.10
- 依赖安装:通过
pip install -r requirements.txt安装所需依赖。

- 1Deconstructing Instruction-Following: A New Benchmark for Granular Evaluation of Large Language Model Instruction Compliance AbilitiesCapital One · 2026年



