Extending LLMs' Context Window with 100 Samples

TL;DR

This paper introduces a novel method to extend LLMs' context window efficiently by adjusting RoPE parameters, validated on LLaMA-2-7B-Chat with minimal data and training, improving performance and robustness.

Contribution

A new extension to RoPE that combines frequency adjustment and attention scaling, enabling large context windows with minimal samples and training steps.

Findings

Extends LLaMA-2-7B-Chat context to 16,384 tokens with only 100 samples.

Improves fine-tuning performance and robustness across different context sizes.

Demonstrates the effectiveness of data composition and training curricula for context extension.

Abstract

Large Language Models (LLMs) are known to have limited extrapolation ability beyond their pre-trained context window, constraining their application in downstream tasks with lengthy inputs. Recent studies have sought to extend LLMs' context window by modifying rotary position embedding (RoPE), a popular position encoding method adopted by well-known LLMs such as LLaMA, PaLM, and GPT-NeoX. However, prior works like Position Interpolation (PI) and YaRN are resource-intensive and lack comparative experiments to assess their applicability. In this work, we identify the inherent need for LLMs' attention entropy (i.e. the information entropy of attention scores) to maintain stability and introduce a novel extension to RoPE which combines adjusting RoPE's base frequency and scaling the attention logits to help LLMs efficiently adapt to a larger context window. We validate the superiority of our method in both fine-tuning performance and robustness across different context window sizes on various context-demanding tasks. Notably, our method extends the context window of LLaMA-2-7B-Chat to 16,384 with only 100 samples and 6 training steps, showcasing extraordinary efficiency. Finally, we also explore how data compositions and training curricula affect context window extension for specific downstream tasks, suggesting fine-tuning LLMs with lengthy conversations as a good starting point. We release our code and SFT data at https://github.com/GAIR-NLP/Entropy-ABF.