ChuckMcSneed/NeoEvalPlusN_benchmark
收藏资源简介:
该数据集是一个用于评估语言模型性能的基准测试集,包含多个测试(B-test、C-test、D-test、P-test、S-test)。B-test、C-test和D-test主要评估模型执行命令的能力,而P-test和S-test则评估模型的创造性写作能力。每个测试都有不同的评分标准和目标。测试结果显示了不同模型的表现及其特点,如模型是否能够通过基本命令测试、创造性写作能力如何等。README还提到了测试的局限性,如测试仅进行一次、人为因素对结果的影响等。
This dataset is a benchmark suite for evaluating language model performance, comprising five test sets: B-test, C-test, D-test, P-test, and S-test. B-test, C-test, and D-test primarily evaluate a model's ability to follow and execute commands, whereas P-test and S-test assess its creative writing proficiency. Each test features distinct scoring criteria and targeted evaluation objectives. The test results demonstrate the performance and characteristics of various models, such as whether a model can pass basic command-following assessments and the quality of its creative writing capabilities. The README also outlines the limitations of these tests, such as the evaluations being conducted only once and the impact of human factors on test outcomes.
数据集概述
数据集目的
- 旨在评估模型在执行命令和创意写作方面的能力,而非智能程度。
测试条件
- 所有测试在koboldcpp环境中执行。
- 测试参数设置:温度(temperature)和顶部概率(top P)约等于0,重复惩罚(rep. penalty)=1。
测试格式
- 使用模型适宜的格式进行测试,除非该格式不适用。




