LONG IS MORE FOR ALIGNMENT

Summary from the TLDR AI newsletter

It is often challenging to know which examples should be used when aligning language models using preference data. This work suggests a surprisingly robust baseline - choose the 1,000 longest examples.