homerquan/mn-context-compression-dataset-v1

该数据集用于训练上下文压缩模型，以提高多智能体运行时（特别是MirrorNeuron及其相关项目）的上下文效率。它用于训练如homerquan/mn-context-engine-lora-v2等模型，以及后续专注于保护事实的上下文引擎。数据强调精确保护跨度保留、源引用保存、预算条件压缩，以及更安全上下文缩减的偏好训练。数据集包括监督微调和偏好训练的分割文件，旨在实现无损上下文压缩，确保操作细节的准确性，例如保留源引用、文件路径、ID、数字、固定事实和最新指令，并根据明确目标预算（如400、800和1200个令牌）条件化输出。它还学习紧凑的保护事实分类账和预算警告，当所需事实超出目标时，优先选择避免私有标记泄漏、过时指令、重复警告和超预算响应的输出。该数据集适用于上下文压缩、内存压缩、SFT、DPO和LoRA适配器的研究和工程工作，特别针对SmolLM3等小型纯文本语言模型，用于评估压缩器是否保留下游智能体行为所需的正确事实，而不仅仅是生成流畅的摘要。This dataset is used to train context-compression models for improving the context efficiency of multi-agent runtimes, especially MirrorNeuron and the broader work at mirrorneuron.io. We use this dataset to train models such as homerquan/mn-context-engine-lora-v2, and later protected-fact-focused context engines. The data emphasizes exact protected-span retention, source-reference preservation, budget-conditioned compression, and preference training for safer context reduction. The dataset includes splits for supervised fine-tuning and preference training, designed for non-lossy context compression where exact operational details matter. It focuses on preserving source refs, file paths, IDs, numbers, pinned facts, and latest instructions; conditioning outputs on explicit target budgets such as 400, 800, and 1200 tokens; learning compact protected-fact ledgers and budget warnings when required facts exceed the target; preferring outputs that avoid private-marker leakage, stale instructions, repeated warnings, and over-budget responses. It is intended for research and engineering work on context compression, memory compaction, SFT, DPO, and LoRA adapters for small text-only language models such as SmolLM3, especially useful for evaluating whether a compressor keeps the facts needed for correct downstream agent behavior, rather than only producing fluent summaries.