File size: 413 Bytes
2e95d53 | 1 2 3 4 5 6 7 8 | # 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)
|