基於可學習編碼簿之大型語言模型提示詞微調

Abstract

提示詞微調是一種熱門的輕量化微調架構，其特點是在預訓練模型的基礎上僅更新少量參數，即可達到亮眼的表現。在過去的方法中，每個提示詞通常被視為一個整體，且各自獨立更新，導致隨著提示詞的增多，更新的參數量也線性增長。為了解決此問題，我們提出了高效率提示詞微調的可適應性編碼簿。我們利用了乘積量化的概念，使提示詞在每個分割後的子空間中共享一組可學習的編碼向量。每個提示詞透過一組自適應權重而有所變化。我們在 17 個自然語言任務中，僅更新預訓練模型 0.3% 的參數，就達到了優異的表現，包括自然語言理解及問答任務。此外，我們的方法在少樣本學習情境以及大型語言模型骨幹下也有良好表現，凸顯了其適應性及可發展性。

Prompt Tuning has emerged as a popular Parameter-Efficient Fine-Tuning method attributed to its excellent performance with few updated parameters on various large-scale Pretrained Language Models (PLMs). In previous approaches, each prompt has been considered as a whole and updated independently, causing all parameters to depend on prompt length and increase accordingly. To alleviate this problem, we introduce Adaptive Codebook for Composite and Efficient Prompt Tuning (ACCEPT). In our approach, we utilize the concept of product quantization (PQ), enabling all soft prompts to share a common set of learnable codebook vectors within each subspace. Each prompt is then distinguished by a unique set of adaptive weights. We achieve impressive performances on 17 diverse natural language tasks, including natural language understanding (NLU) and question answering (QA), by training only 0.3% of parameters of the PLMs. Additionally, our method excels in fewshot and large model scenarios, highlighting its significant adaptability and potential.