baitao666/Step-3.5-Flash-SFT

Step-3.5-Flash-SFT 是一个用于聊天模型的通用领域监督微调数据集，支持多语言文本生成任务，专注于聊天、指令调优、推理和代码代理。数据集以JSON格式组织，包含用户与助手之间的对话轮次，每个轮次包含角色、内容、损失掩码等字段，部分助手响应还包含推理内容。它设计用于StepTronOSS训练框架，需遵循特定训练语义（如顺序采样、损失掩码应用规则），并提供了分词器快照和编译后的分片以加速训练。数据集强调作为训练语料库而非基准测试，使用时需同时遵守Apache-2.0和CC-BY-NC-2.0许可证。Step-3.5-Flash-SFT is a general-domain supervised fine-tuning dataset for chat models, supporting multilingual text generation tasks, with a focus on chat, instruction-tuning, reasoning, and code agent applications. The dataset is organized in JSON format, containing conversation turns between users and assistants, each with fields such as role, content, loss mask, and optional reasoning content for assistant responses. It is designed for use with the StepTronOSS training framework, requiring adherence to specific training semantics (e.g., sequential sampling, loss mask rules), and includes tokenizer snapshots and compiled shards for accelerated training. The dataset is intended as a training corpus rather than a benchmark, and usage must comply with both the Apache-2.0 and CC-BY-NC-2.0 licenses simultaneously.