FINAL-Bench/Aether-7B-5Attn-checkpoints
收藏资源简介:
Aether-7B-5Attn中间预训练检查点数据集,包含在预训练过程中生成的三个中间检查点(步骤110,000、115,000和162,000),用于支持完全可重复性和训练动态研究,符合OLMo级别的“完全开放”标准。这些检查点对应于近似98B、102B和144B令牌的训练数据,其中162,000步骤的检查点是最终退火基准模型的原始版本。每个检查点文件夹包括model.safetensors、config.json和tokenizer文件,基于自定义架构aether_v2_7way,需使用基础模型仓库中提供的aether_pkg代码加载,并在推理时以batch_size=1运行。
Aether-7B-5Attn Intermediate Pre-training Checkpoint Dataset contains three intermediate checkpoints generated during pre-training (steps 110,000, 115,000, and 162,000). This dataset is designed to support full reproducibility and training dynamics research, and complies with the "fully open" standard at the OLMo level. These checkpoints correspond to training data of approximately 98B, 102B, and 144B tokens respectively, and the checkpoint at step 162,000 is the original version of the final annealed baseline model. Each checkpoint folder includes model.safetensors, config.json, and tokenizer files. Based on the custom architecture aether_v2_7way, these checkpoints require the aether_pkg code provided in the base model repository for loading, and must be run with a batch_size=1 during inference.




