alex-karev/llm-comparison

该数据集基于原始的Alpaca数据集，使用ConSCompF框架生成，用于LLM相似性比较。数据集分为三个子集：quantization包含156,000个样本，由Tinyllama及其8位、4位和2位GGUF量化版本生成；comparison包含28,600个样本，由Tinyllama、Mistral、OpenHermes2.5、Gemma2、Qwen2.5、Phi3.5-mini、GLM4和Llam2的4位量化版本生成；prompts包含39,000个样本，由Gemma2-2b、Phi3.5-mini和Qwen2.5-3b的4位量化版本生成，并使用了5种不同的系统提示设置。

This dataset is based on the original Alpaca dataset and was synthetically generated for LLM similarity comparison using the ConSCompF framework. It is divided into three subsets: quantization contains 156,000 samples generated by Tinyllama and its 8-bit, 4-bit, and 2-bit GGUF quantized versions; comparison contains 28,600 samples generated by 4-bit quantized versions of Tinyllama, Mistral, OpenHermes2.5, Gemma2, Qwen2.5, Phi3.5-mini, GLM4, and Llam2; prompts contains 39,000 samples generated by 4-bit quantized versions of Gemma2-2b, Phi3.5-mini, and Qwen2.5-3b, with 5 different system prompt setups.