a9lim/llmoji

llmoji数据集是一个由贡献者提交的颜文字（kaomoji）及其含义的语料库。该数据集旨在探索模型在回复开头选择颜文字是否反映其内部状态。数据集包含JSONL文件，记录颜文字、使用次数以及对其含义的综合描述。数据收集过程涉及从模型回复中提取颜文字的钩子，数据集结构包括清单文件和每个源模型的JSONL文件。数据集采用CC-BY-SA-4.0许可，专为研究目的设计，特别是在机械可解释性和LLM行为研究方面。

The llmoji dataset is a contributor-submitted corpus of kaomoji used by AI models and their meanings. The dataset is designed to explore whether a models choice of kaomoji at the start of a reply reflects its internal state. It includes JSONL files with kaomoji, their usage counts, and synthesized descriptions of their meanings. The data collection process involves hooks that extract kaomoji from model replies, and the dataset is structured with manifest files and per-source-model JSONL files. The dataset is licensed under CC-BY-SA-4.0 and is intended for research purposes, particularly in mechanistic interpretability and LLM behavior studies.