Tioe/LATENT-SWITCH-69K

LATENT-SWITCH-69K是一个用于潜在推理训练的数据集，包含69,745个经过处理的监督微调样本。数据集采用Parquet格式，主要文件为sft_train.parquet，包含31个列，于2026年4月生成。数据源预处理模式为sft，使用Hugging Face分词器进行计数，参考分词器路径为https://huggingface.co/Qwen/Qwen3-14B。数据集包括提示/响应消息、蒸馏的思维链文本、潜在推理元数据、排序字段和状态对齐监督字段。主要字段有记录ID、问题、真实答案、消息、助手思维链、助手答案、难度、潜在步骤数等。数据集难度分布为简单6,667行、中等45,650行、困难17,428行，潜在步骤数范围为5到128。该数据集旨在用于潜在推理的监督微调实验，但用户在使用前应检查示例和元数据。LATENT-SWITCH-69K is a dataset for latent reasoning training, containing 69,745 processed supervised fine-tuning samples. The dataset is in Parquet format, with the main file being sft_train.parquet, consisting of 31 columns, generated in April 2026. The source preprocessing mode is sft, using the Hugging Face tokenizer for counting, with the reference tokenizer path being https://huggingface.co/Qwen/Qwen3-14B. It includes prompt/response messages, distilled chain-of-thought text, latent reasoning metadata, ordering fields, and state-alignment supervision fields. Key fields include record ID, question, ground truth, messages, assistant chain-of-thought, assistant answer, difficulty, number of latent steps, etc. The difficulty distribution is 6,667 easy rows, 45,650 medium rows, and 17,428 hard rows, with latent steps ranging from 5 to 128. This dataset is intended for supervised fine-tuning experiments in latent reasoning, but users should review examples and metadata before use.