Kukedlc/omnicoder-train

omnicoder-train数据集包含16178个预格式化示例，使用Tesslate/OmniCoder-9B的官方模板，基于Qwen3.5-9B模型。数据经过过滤，确保包含结构化思考标记<think>...</think>，总token数少于2048，并涵盖数学、代码、逻辑、编程、物理和化学等类别。数据来自多个公开源，如roman、jirafa、nohurry等，格式为JSONL，包含文本、token数、来源、类别和过滤标签等字段。适用于监督微调训练，特别是与HuggingFace的SFTTrainer结合使用。

The omnicoder-train dataset contains 16,178 pre-formatted examples using the official template of Tesslate/OmniCoder-9B, based on Qwen3.5-9B. The data is filtered to include structured thinking markers <think>...</think>, have less than 2048 tokens, and cover categories such as math, code, logic, coding, physics, and chemistry. It is sourced from multiple public datasets like roman, jirafa, nohurry, etc., in JSONL format with fields including text, n_tokens, source, category, and filter_label. It is suitable for supervised fine-tuning training, particularly with HuggingFaces SFTTrainer.