This commit adds the trained logistic regression (LR) probing model files used in our paper:
**"Sentinel: Attention Probing of Proxy Models for LLM Context Compression with an Understanding Perspective"**
These models perform sentence-level relevance prediction based on decoder attention features from a 0.5B-scale proxy model (Qwen2.5-0.5B-Instruct). The logistic regression classifier is trained on attention-derived features from SQuAD, NewsQA, and HotpotQA under weak supervision.
📂 Files:
- `.pkl`: Serialized LR classifier weights
- `.json`: Feature configuration used during probing
These files correspond to the setup described in Section 2.3 of the paper. Use them to replicate the results in our LongBench evaluations.
📄 Paper: https://arxiv.org/abs/2505.23277
🔗 Code: https://github.com/yzhangchuck/Sentinel
