syntropicsignal-ai/wildchat-asking-en-text-embedding-3-small

WildChat Asking-mode (EN) — text-embedding-3-small 是一个从allenai/WildChat-1M数据集中提取的英文首轮用户提示数据集，经过过滤仅保留Asking-mode（询问模式）提示，并使用OpenAI的text-embedding-3-small模型进行嵌入处理。数据集包含189,916行数据，语言为英文，嵌入维度为1536（经过L2归一化），格式为单个Parquet文件（采用zstd压缩）。许可证为ODC-BY（继承自WildChat-1M）。构建过程包括源数据流处理、基于前缀的保守分类器过滤询问模式提示，以及通过OpenAI Embeddings API进行嵌入。预期用于检索、相似性搜索、分类和提示研究等任务，特别适合代表用户向助手询问的内容（而非用户要求助手生成的内容）。局限性包括启发式过滤器可能误分类、依赖单一嵌入模型、领域偏差（反映2023-2024年ChatGPT使用情况）以及无额外PII保证。WildChat Asking-mode (EN) — text-embedding-3-small is a dataset of English-language first-turn user prompts sourced from allenai/WildChat-1M, filtered to retain only Asking-mode prompts and embedded with OpenAIs text-embedding-3-small model. It contains 189,916 rows, with English language, an embedding dimension of 1536 (L2-normalized), and is provided as a single Parquet file with zstd compression. The license is ODC-BY (inherited from WildChat-1M). The dataset was built by streaming the source, applying a conservative prefix-based classifier to filter out Doing-mode prompts, and embedding each surviving prompt via the OpenAI Embeddings API. It is intended for retrieval, similarity search, classification, and prompt-research workloads, particularly suited to representing what users ask assistants rather than what they ask assistants to generate. Limitations include heuristic filter inaccuracies, dependency on a single embedding model, domain skew reflecting ChatGPT usage in 2023-2024, and no additional PII guarantees beyond upstream handling.