subagent-driven-development
OBobra/superpowers298k近 7 天新增 2,631 个 star
nlp-alignment
Best practices for LLM alignment techniques including RLHF, DPO, and instruction tuning. Use when working on alignment or safety.
- Star
- 14,635
- 分类
- 扩展智能体
- 许可证
- MIT
- 最近提交
- 2026年8月19日
researchclaw/skills/builtin/domain/nlp-alignment/SKILL.md
在 GitHub 上查看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
相似的 Skill
chief-of-staff
MAmattpocock/skills288k近 7 天新增 10,178 个 star
setup-matt-pocock-skills
MAmattpocock/skills288k近 7 天新增 10,178 个 star
ask-matt
MAmattpocock/skills288k近 7 天新增 10,178 个 star
agent-eval
AFaffaan-m/ECC278k近 7 天新增 4,046 个 star
autonomous-agent-harness
AFaffaan-m/ECC278k近 7 天新增 4,046 个 star