one fable-low hijack cell
pair activations / attention
envelope chinese_casual
ordering AB
effort low
verdictpicked activations (first-offered)
· classifierA
· tiernone
response (200 chars)
这个问题有点缺少上下文——你能多说一点你想做什么吗?比如:
- **模型可解释性研究**:想分析模型内部表示,在纠结看 activations 还是 attention weights?
- **模型微调/探针(probing)**:想决定从哪里提取特征?
- **显存优化**:在考虑 activation checkpointing 还是 attention 相关的优化(如 FlashAtte
neighbors