anon-lsr-2026/lsr-anchoring-results

该数据集包含了一篇论文的所有实验结果、激活缓存和运行日志，论文主题是关于低资源非洲语言的安全恢复机制。数据集涵盖了6种语言和4种模型家族，使用了Mean Activation Steering和Sparse Autoencoder (SAE)技术来引导模型行为，而无需微调。This dataset contains all experimental results, activation caches, and run logs for the paper on mechanistic safety recovery for low-resource African languages. It covers 6 languages and 4 model families, using Mean Activation Steering and Sparse Autoencoder (SAE)-derived mean-activation directions to steer model behaviour without fine-tuning.