TheFinAI/dolma3_300B_sample_shuffled
收藏资源简介:
dolma3_300B_sample_shuffled数据集是对TheFinAI/dolma3_300B_sample数据集的全局行级洗牌版本。源数据来自allenai/dolma3_mix-6T-1025-7B,通过伯努利采样(概率约0.0506)生成约3000亿cl100k令牌,保留了原始Dolma3的混合比例。由于源数据在磁盘上按子源聚类存储,可能导致训练时小缓冲区出现非均匀源混合问题。本数据集通过真正的全局洗牌解决了这一问题:每行数据均匀随机分配到200个输出桶中,每个桶在内存中进行洗牌,同时保持源混合比例不变。数据模式与源相同,包括来源、日期、文本、令牌计数和类别字段。洗牌使用种子42以确保可重复性。
The dolma3_300B_sample_shuffled dataset is a globally row-shuffled variant of the TheFinAI/dolma3_300B_sample dataset. Its source data originates from allenai/dolma3_mix-6T-1025-7B, and approximately 300 billion cl100k tokens were generated via Bernoulli sampling with an approximate probability of 0.0506, while preserving the original mixture ratios of Dolma3. Since the source data is stored on disk clustered by sub-sources, it may introduce non-uniform source mixing issues in small buffers during training. This dataset addresses this problem via true global shuffling: each row of data is uniformly randomly assigned to one of 200 output buckets, each bucket is shuffled in-memory, and the source mixture ratios are maintained unchanged. The data schema is identical to that of the source dataset, including fields such as source, date, text, token count, and category. The shuffling process uses a fixed seed of 42 to ensure reproducibility.




