marin-community/grug-moe-mix-swarm
收藏资源简介:
Grug-MoE Fisher-DSP Mix Swarm (840 runs) 是一个包含840个MoE预训练运行的数据集,源自Grug-MoE Fisher-DSP数据混合群(维度512,运行于TPU/us-central2)。每个运行在168个数据桶的不同数据混合上进行训练;该群用于回归混合权重→评估损失,并预测优化的预训练混合。数据集包含实验索引、阶段0和阶段1的混合权重(字典格式,总和为1)、模型路径、评估任务数量、宏观比特每字节(bpb)以及每个任务的评估结果(以JSON字符串存储)。数据仅包含元数据(混合权重、评估损失、检查点路径字符串),不包含模型权重本身。
Grug-MoE Fisher-DSP Mix Swarm (840 runs) is a dataset containing 840 MoE pre-training runs, derived from the Grug-MoE Fisher-DSP data mixture swarm with a dimension of 512, deployed on TPU/us-central2. Each run is trained on distinct data mixtures across 168 data buckets; this dataset is designed to regress mixture weights, evaluate corresponding losses, and predict optimized pre-training mixtures. The dataset includes experimental indices, mixture weights for Stage 0 and Stage 1 (formatted as a dictionary with a sum of 1), model paths, the number of evaluation tasks, macro bits per byte (bpb), and per-task evaluation results stored as JSON strings. The dataset only contains metadata including mixture weights, evaluation losses, and checkpoint path strings, and does not include the model weights themselves.




