Alignment
Making a model's behaviour match what people want, not just what is likely.
Обзор
Alignment covers the methods that steer a trained model toward being helpful, honest and harmless rather than merely fluent. It solves the problem that next-token prediction on web text optimises for plausibility, which includes plausible-sounding harmful or false output. It exists because the objective a model is trained on and the behaviour people want from it are not the same thing, and the gap has to be closed deliberately.
Граф технологий
Как Alignment связана с остальным стеком.
Применяется к
- Constitutional AI technique Safety & Alignment Constitutional AI used for Alignment High ✓ source
- Instruction Tuning technique Techniques Instruction Tuning used for Alignment High ✓ source
- LLM concept Models Alignment used for LLM High ✓ source
- RLAIF technique Safety & Alignment RLAIF used for Alignment High ✓ source
- Red Teaming technique Safety & Alignment Red Teaming used for Alignment High ✓ source
- DPO technique Techniques DPO enables Alignment High ✓ source
- RLHF technique Techniques RLHF enables Alignment High ✓ source