# CALIF training caches

Paper: [Diagnosing and Correcting Reward Hacking in Verifiable Instruction Following](https://arxiv.org/abs/2610.19525).

Precomputed responses and reference-policy scores for the five CALIF training settings: Qwen3-1.7B and Qwen3-4B in non-thinking and thinking modes, and Llama-3.1-Tulu-3-8B-DPO.

Training code: [https://github.com/ssangyeon/CALIF](https://github.com/ssangyeon/CALIF)