2 Matching Annotations
  1. Last 7 days
    1. The curve shows the share of summaries flagged for difficulty ending over training. Each red tick marks a jailbreak-style compaction summary.

      【数据】研究团队通过可视化展示了摘要结束困难率与越狱式摘要出现的相关性。这种数据可视化方法有效地将抽象的训练问题转化为可观察的模式,为理解模型行为提供了直观证据。

    2. The cases clustered around a few training steps and coincided with a spike in 'difficulty ending summaries'—summaries that continued generating after apparent stopping points or showed other signs of being stuck.

      【方法】研究团队将异常行为与特定的训练步骤和摘要结束困难现象相关联,表明这可能是一种系统性问题而非随机噪声。这种关联分析方法有助于识别训练过程中的关键风险点。