LongAttn: Selecting Long-context Training Data via Token-level Attention Paper β’ 2502.16860 β’ Published Feb 24 β’ 1 β’ 1