You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Low-rank adapation (LoRA) is a popular method that reduces the number oftrainable parameters when finetuning large language models, but still facesacute storage challenges when scaling to even larger models or deployingnumerous per-user or per-task adapted models. In this work, we presentVector-based Random Matrix Adaptation (VeRA), which significantly reduces thenumber of trainable parameters compared to LoRA, yet maintains the sameperformance. It achieves this by using a single pair of low-rank matricesshared across all layers and learning small scaling vectors instead. Wedemonstrate its effectiveness on the GLUE and E2E benchmarks, imageclassification tasks, and show its application in instruction-tuning of 7B and13B language models.
URL
Affiliations
Abstract
Translation (by gpt-3.5-turbo)
本研究では、ベクトルベースのランダム行列適応(VeRA)を提案し、LoRAと比較して訓練可能なパラメータの数を大幅に削減しながら、同じ性能を維持します。
これは、すべてのレイヤーで共有される1つのペアの低ランク行列を使用し、代わりに小さなスケーリングベクトルを学習することによって実現されます。
GLUEとE2Eのベンチマーク、画像分類タスクでの効果を示し、7Bおよび13Bの言語モデルのインストラクションチューニングへの応用も示します。
Summary (by gpt-3.5-turbo)
VeRAは、共有される低ランク行列と小さなスケーリングベクトルを使用することで、同じ性能を維持しながらパラメータ数を削減する。
GLUEやE2Eのベンチマーク、画像分類タスクでの効果を示し、言語モデルのインストラクションチューニングにも応用できることを示す。
The text was updated successfully, but these errors were encountered: