enuno LLM Alignment Best Practice
Methods:
- RLHF: Train reward model → PPO fine-tuning (complex but powerful)
- DPO: Direct preference optimization (simpler, no reward model needed)
- GRPO: Group relative policy optimization
- SFT: Supervised fine-tuning as alignment baseline
Training recipe:
- Start with SFT on high-quality instruction data
- DPO: lr=5e-7, beta=0.1, batch_size=64
- PPO: lr=1e-6, clip=0.2, KL coeff=0.02
- Use reference model for KL penalty
- Evaluate on safety benchmarks (TruthfulQA, BBQ, etc.)
Common pitfalls:
- Reward hacking: model finds shortcuts to high reward
- Mode collapse: model generates repetitive outputs
- Catastrophic forgetting: loses general capabilities