karlexmarin/taf-attention-decay
收藏资源简介:
TAF注意力衰减测量数据集是第一个公开的关于transformer LLMs注意力衰减指数γ测量的数据集。该数据集包含35个模型,13个家族,88条记录,覆盖了两种语料库(真实文本和随机标记)和两种精度(4-bit NF4和bfloat16)。数据集还包括随机初始化控制、Pythia-70M训练轨迹、Pythia-31m高n鲁棒性、Yi-9B随机标记等特定内容。注意力衰减指数γ是一个单数诊断指标,用于衡量transformer的注意力是局部还是全局的,它连接了RoPE几何与长上下文行为、KV缓存压缩、NIAH检索和幻觉率等。
The TAF Attention-Decay Measurements dataset is the first public dataset of attention-decay exponent γ measurements across transformer LLMs. It includes 35 models from 13 families, totaling 88 records, covering two corpora (real text and random tokens) and two precisions (4-bit NF4 and bfloat16). The dataset also features random-init controls, Pythia-70M training trajectory, Pythia-31m high-n robustness, and Yi-9B random_tokens, among others. The attention-decay exponent γ is a single-number diagnostic of how locally or globally a transformer attends, connecting RoPE geometry to long-context behavior, KV-cache compression, NIAH retrieval, and hallucination rates.




