Together AI 长上下文微调：突破 LLM 有效长度限制
原标题：The landscape of Large Language Models (LLMs) is rapidly evolving, with context lengths expanding from a few thousand tokens a year ago to millions of tokens now. This increase in context length has v
AI 摘要
Together AI 发布技术博客，探讨大语言模型长上下文能力的现状与挑战。文章指出，尽管模型宣称支持百万级 token，但实际有效上下文长度远低于标称值，性能随长度增加而下降。为解决此问题，Together AI 平台现已支持最长 32k token 的微调，并展示了通过微调 Llama 3.1 8B 模型在重复任务上性能大幅提升的案例，强调长上下文微调对企业的价值。
正文节选
The landscape of Large Language Models (LLMs) is rapidly evolving, with context lengths expanding from a few thousand tokens a year ago to millions of tokens now. This increase in context length has very real implications for enterprise applications, particularly in Retrieval Augmented Generation (RAG), document analysis, and summarization systems. While prior models were limited to processing a few pages of text, modern models like Meta's Llama 3.2 series can handle 131K tokens, which is the eq