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