Paper summary
Paper summary: Self-Alignment with Instruction Backtranslation
Instruction backtranslation lets a language model label unlabelled web text with the instructions it answers, then pick the best pairs to fine-tune itself.
- The model writes the missing instruction for an existing human-written answer, which is the reverse of normal generation.
- Self-curation, where the model scores its own candidate pairs, matters as much as self-augmentation.