content of 'labels' when doing prompt tuning of llama-2 on QA

bigscience-workshop / petals

🌸 Run LLMs at home, BitTorrent-style. Fine-tuning and inference up to 10x faster than offloading

https://petals.dev

MIT License

9.26k stars 524 forks source link

content of 'labels' when doing prompt tuning of llama-2 on QA #561

Open caochuxueeee opened 8 months ago

caochuxueeee commented 8 months ago

I saw that on classification tasks, the 'labels' are target values. When using CausalLM model to tune it on QA dataset, the format should be 1) input_ids: Q; labels:A or 2) input_ids: Q+A, labels: Q+A or 3) input_ids: Q+A, labels: IGNORE_TOKENs+A ?

Thank you!