Attention Alignment and Flexible Positional Embeddings Improve Transformer Length Extrapolation

Re-assign community