ceselder/loracle-ptrl-data-v9
收藏资源简介:
该数据集包含了复现LoRA Oracles项目中v9 keyword-judge RL训练运行所需的一切内容。loracle通过读取LoRA的权重(而非输出)来预测LoRA的行为。v9测试了一个特定假设:强制loracle预测主题/关键词(而非完整文档文本)是否能恢复表面文档匹配法官错过的隐藏行为。结果显示,仅经过10个RL周期后,loracle能够成功预测所有四个动物的隐藏特征,而之前只能预测鲸鱼。数据集包含RL数据、关键词列表、训练配置和法官提示等文件。
This dataset contains everything needed to reproduce the v9 keyword-judge RL training run for the LoRA Oracles project. A loracle reads a LoRAs weights (not its outputs) and predicts what the LoRA does. v9 tests a specific hypothesis: does forcing the loracle to predict themes/keywords (instead of full document text) recover hidden behaviors that surface-document-matching judges missed? The headline result, after only 10 RL cycles, is that the loracle successfully predicted all four animals hidden traits, whereas previously it could only predict whales. The dataset includes RL data, keyword lists, training configurations, and judge prompts.




