CALIF training caches
Paper: Diagnosing and Correcting Reward Hacking in Verifiable Instruction Following.
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