68K (instruction, input, output) instances in MUFFIN
IDEA
EXECUTION
Image
Instruction extraction
given a piece of text, get an LLM to suggest different textual facets of the text ⇒ use those facets to suggest possible instructions (+ three shot from instructions from SuperNI)
Instruction rematching
gather suitable instructions from existing human-annotated set ⇒ ask GPT-4 whether the instruction can align with the input to form a valid task (“Yes” or “No”).
Creating outputs
Given the created (instruction, input pairs), use ChatGPT to annotate outputs, or mark them as “None” for instances that ChatGPT deems unanswerable