search

YJiangcm / Lion

Code for "Lion: Adversarial Distillation of Proprietary Large Language Models (EMNLP 2023)"
https://arxiv.org/abs/2305.12870
MIT License
201 stars 19 forks source link

adout second iteration and new_instruction #9

Open Git-L1 opened 7 months ago

Git-L1 commented 7 months ago

hi, thanks for release your code.I have some small questions to ask you. The code you provided is for one iteration, so what is the input dataset for the second iteration? Is it the train_cool that has been replaced in the diagram, and where is cache_pool reflected in the code? Then,Does the new_instruction generated in the diagram refer to the merger of new_hard and new_easy? Looking forward to your early reply. thanks! QQ图片20240330212127

YJiangcm commented 7 months ago

Thanks for your interest in our work. Let me illustrate it using an example:

In the first iteration, the _trainpool and _cachepool are both 52,000 Alpaca instructions and we generate 3,000 _newhard and 3,000 _neweasy (the combination of them is regarded as _newinstruction).

Therefore, in the second iteration, the _trainpool would be exactly the 6,000 _newinstruction. The _cachepool would be the instructions from the previous iteration plus the _newinstruction, which are 58,000 instructions in total.