datumo/Feedback-Bench
收藏资源简介:
BENCHMARK-Feedback-Bench-v0.1 是一个用于测量基于量规的评估能力(rubric following)的基准数据集。它旨在比较评估器在韩语环境(Ko Feedback Bench)和英语环境(Feedback Bench)下的性能,并用于验证所提出评估器的量规遵循能力。数据集总共有2000个实例,平均分为两个子集:ko_feedback_bench(1000个韩语实例)和feedback_bench(1000个英语实例)。每个实例使用5点量表(1-5分)进行标注。数据集基于原始的Feedback Bench(Kim et al., 2024a)构建,但采用了无参考(reference-free)设置和基于辩论(debate)的重新标注过程,以更真实地反映工业评估环境并严格测量评估器的独立量规遵循能力。韩语子集是通过将原始英语的指令、量规和模型响应翻译成韩语并经过审核后构建的,而英语子集则是在移除参考答案后通过相同的辩论过程重新标注的。数据集包含多个特征字段,如原始指令、评估标准、各分数点的量规描述、模型响应、用户提示、系统提示和原始分数。需要注意的是,韩语版本可能存在机器翻译带来的噪声,且数据集的标签是基于共识的,更多反映与共识标注标准的一致性而非绝对评估质量。
BENCHMARK-Feedback-Bench-v0.1 is a benchmark dataset designed to measure rubric-following evaluation capabilities. It is used to compare the performance of evaluators in Korean (Ko Feedback Bench) and English (Feedback Bench) environments, and to validate the rubric-following ability of proposed evaluators. The dataset contains a total of 2,000 instances, split evenly into two subsets: ko_feedback_bench (1,000 Korean instances) and feedback_bench (1,000 English instances). Each instance is labeled on a 5-point scale (1–5). The dataset is based on the original Feedback Bench (Kim et al., 2024a) but employs a reference-free setting and a debate-based relabeling process to better reflect real-world industrial evaluation scenarios and rigorously measure the evaluators independent rubric-following ability. The Korean subset was constructed by translating the original English instructions, rubrics, and model responses into Korean, followed by review, while the English subset was reconstructed by removing reference answers and relabeling through the same debate process. Key features include original instruction, evaluation criteria, rubric descriptions for each score point, model response, user prompt, system prompt, and original score. It should be noted that the Korean version may contain noise due to machine translation, and the dataset labels are consensus-based, reflecting alignment with consensus labeling standards rather than absolute evaluation quality.




