CryptoLLM — Cryptography Foundation Model
Summary
World's first cryptography LLM. Led the full CPT→SFT→DPO→GRPO post-training pipeline, outperforming GPT-4o by 8.8% on CryptoBench.
Problem
General LLMs are weak in cryptography's term-dense, rigorous reasoning across algorithms, protocols and engineering, with no trustworthy domain benchmark.
Solution
Domain-adapted Qwen2.5-72B: multi-source CPT (~5B tokens), self-instruct SFT/DPO data, and GRPO with verifiable rewards; built a 2K-item CryptoBench with difficulty stratification.
Impact
- Outperformed GPT-4o by 8.8% on CryptoBench with no meaningful loss on C-Eval / MMLU.
- Established a reusable domain post-training pipeline and trustworthy evaluation standard.
Post-training Pipeline
CPT (multi-source cleaning/dedup/replay) → SFT (loss mask + packing) → DPO (rejection-sampled pairs) → GRPO (verifiable rewards, 8×H100, verl).
Key Techniques
Key Insight · Silent Groups
In GRPO, all-correct/all-wrong groups have zero reward variance and zero advantage; difficulty filtering keeps mid-difficulty items — the same principle as evaluation discrimination.