THU-KEG/IFBench
收藏Hugging Face2025-03-07 更新2025-04-12 收录
下载链接：
https://hf-mirror.com/datasets/THU-KEG/IFBench
下载链接
链接失效反馈官方服务：
资源简介：
IFBench是一个用于评估指令遵循奖励模型的基准数据集。它包含了唯一标识符、源数据集、原始指令和增强指令、选择的响应和拒绝的响应，以及需要基于LLM或基于代码验证的约束。该数据集与论文《Agentic Reward Modeling: Integrating Human Preferences with Verifiable Correctness Signals for Reliable Reward Systems》相关。
IFBench is a benchmark dataset for evaluating instruction-following reward models. It includes a unique identifier, source dataset, original and augmented instructions, chosen and rejected responses, and constraints that require LLM-based or code-based verification. The dataset is related to the paper Agentic Reward Modeling: Integrating Human Preferences with Verifiable Correctness Signals for Reliable Reward Systems.
提供机构：
THU-KEG
