LSG Attention: Extrapolation of pretrained Transformers to long sequences
Paper • 2210.15497 • Published • 1
LSG16K-Italian-LEGAL-BERT
Local-Sparse-Global version of ITALIAN-LEGAL-BERT by replacing the full attention in the encoder part using the LSG converter script (https://github.com/ccdv-ai/convert\_checkpoint\_to\_lsg). We used the LSG attention with 16,384 maximum sequence length, 7 global tokens, 128 local block size, 128 sparse block size, 2 sparsity factors, 'norm' sparse selection pattern (select the highest norm tokens).
- Downloads last month
- 338