david-thrower/HelixLM-tiny-400.0Mt-730000pt-57143it-20260430
收藏资源简介:
该数据集是一个经过缩放的子集,从三个高质量来源中精选而来,推荐比例为:FineWeb-Edu(约85%)、OpenWebMath(约10%)和OpenHermes-2.5(约5%)。数据集分为预训练和指令调优两部分:预训练部分包括pretrain_train和pretrain_val(来自FineWeb-Edu和OpenWebMath),指令调优部分包括instruct_train和instruct_val(来自OpenHermes-2.5,格式化为包含系统、用户、助手和结束标记的对话形式)。数据准备包括流式加载、10,000的随机缓冲区、2%的验证集保留以及鲁棒的指令格式检测。
This dataset is a scaled subset curated from three high-quality sources in the recommended ratios: FineWeb-Edu (~85%), OpenWebMath (~10%), and OpenHermes-2.5 (~5%). The dataset is split into pretraining and instruction tuning portions: the pretraining portion includes pretrain_train and pretrain_val (from FineWeb-Edu and OpenWebMath), and the instruction tuning portion includes instruct_train and instruct_val (from OpenHermes-2.5, formatted with system, user, assistant, and end-of-text markers). Data preparation involves streaming load with a shuffle buffer of 10,000, a 2% validation holdout per split, and robust schema detection for instruction formatting.




