Long Context Alignment with Short Instructions and Synthesized Positions
This addresses the problem of long-context processing for LLM users by offering an efficient alignment method, though it appears incremental as it builds on existing techniques without a paradigm shift.
The paper tackles the challenge of enabling large language models to handle long-context instructions without requiring high-quality long data or extra computational resources, and introduces Step-Skipping Alignment (SkipAlign), which achieves performance comparable to GPT-3.5-Turbo-16K on LongBench with only 6B parameters.
Effectively handling instructions with extremely long context remains a challenge for Large Language Models (LLMs), typically necessitating high-quality long data and substantial computational resources. This paper introduces Step-Skipping Alignment (SkipAlign), a new technique designed to enhance the long-context capabilities of LLMs in the phase of alignment without the need for additional efforts beyond training with original data length. SkipAlign is developed on the premise that long-range dependencies are fundamental to enhancing an LLM's capacity of long context. Departing from merely expanding the length of input samples, SkipAlign synthesizes long-range dependencies from the aspect of positions indices. This is achieved by the strategic insertion of skipped positions within instruction-following samples, which utilizes the semantic structure of the data to effectively expand the context. Through extensive experiments on base models with a variety of context window sizes, SkipAlign demonstrates its effectiveness across a spectrum of long-context tasks. Particularly noteworthy is that with a careful selection of the base model and alignment datasets, SkipAlign with only 6B parameters achieves it's best performance and comparable with strong baselines like GPT-3.5-Turbo-16K on LongBench.