issai/foggen-data

FogGen数据集是一个用于复现FogGen模型（基于Qwen/Qwen3-0.6B训练）的评估和训练数据集合。数据集包含以下内容：1) SFT（监督微调）输入数据，用于训练R14（14轮训练后的FogGen链端点）和R15-OE扩展模型，共约3600行对话格式数据，其中助理响应包含置信度和最终答案。2) 逐问题评估输出，包括多项选择题（MCQ）和开放性问题（OE）的评估结果，覆盖7个领域：金融、科学、编码、法律、数学、哈萨克文化和医学，每个领域包含完整测试集（约16k查询）的N=8自验证样本。3) 对比数据，包括原始Qwen/Qwen3-0.6B模型（未经FogGen链训练）的评估输出，以及云基线输出。数据格式为JSONL，包含问题、选项、黄金答案、验证结果、置信度、评分等字段，用于分析边缘云路由、自信度表达和模型性能提升。数据集支持通过Hugging Face datasets库加载，并包含汇总统计信息，用于复现论文中的主要结果。The FogGen Dataset is an evaluation and training dataset for reproducing the FogGen model, which is trained based on Qwen/Qwen3-0.6B. The dataset includes the following contents: 1) SFT (Supervised Fine-tuning) input data, used for training R14 (the FogGen chain endpoint after 14 training rounds) and R15-OE extended models, with approximately 3600 dialogue-format data entries, where the assistant responses contain confidence scores and final answers. 2) Per-question evaluation outputs, including evaluation results for multiple-choice questions (MCQ) and open-ended questions (OE), covering 7 domains: finance, science, coding, law, mathematics, Kazakh culture, and medicine. Each domain includes N=8 self-verification samples for the full test set (approximately 16k queries). 3) Comparative data, including evaluation outputs of the original Qwen/Qwen3-0.6B model (untrained via the FogGen chain) and cloud baseline outputs. The dataset is stored in JSONL format, with fields such as question, options, golden answer, verification results, confidence, scores, etc., which are used for analyzing edge-cloud routing, confidence expression and model performance improvement. The dataset supports loading via the Hugging Face datasets library, and includes summary statistics for reproducing the main results in the associated paper.