Training on logits rather than tokens?

artidoro / qlora

QLoRA: Efficient Finetuning of Quantized LLMs

https://arxiv.org/abs/2305.14314

MIT License

9.74k stars 800 forks source link

Training on logits rather than tokens? #271

Open SinanAkkoyun opened 9 months ago

SinanAkkoyun commented 9 months ago

Hey, I would like to train a student model from my teacher model (knowledge distillation for specualtive decoding). Commonly, the student model is being trained on the teachers logits (soft "labels") rather than tokens (hard "labels")

How can I do that with qLora? Thank you so much!