Datasets:

mii-llm

/

gazzetta-ufficiale

like 25

Hi. My approach would be to create a synthetic question from a given context (choosing a type, e.g., "true/false", "summary", "open-ended question", "multiple choice", etc.) and then, using a strong LLM, answer the question using the context. Then you have question-answer tuples, so you can train an LLM. You can also combine this approach with different ones:

• use text as output and create a synthetic question & summary: ("Scrivi l'atto ... a partire dalla descrizione: [synthetic summary]", "[text]")
• classification: ("Classifica l'atto ... fra: ...", "[field or subfield]")
  ...

Thanks, Edoardo 😄️