We introduce goldfish loss, a new language modeling loss function that mitigates memorization of training data.
Specifically, goldfish loss pseudorandomly drops $1/k$ of total tokens seen (in the forward pass) during loss computation (i.e., it doesn't compute loss for these tokens), with k being a hyperparameter.
We show that the model finds it increasingly difficult to verbatim regurgitate training data even after 100 epochs. Please read our paper linked below for more details.
The following checkpoints are from our paper titled Goldfish Loss: Mitigating Memorization in Generative LLMs [
paper link].
Each checkpoint mentioned above used randomly initialized
TinyLLaMA-1.1B architecture.
For pretraining details, please find check our
GitHub repository.
If you find our model, codebase or dataset beneficial, please consider citing our work:
1@misc{hans2024like,
2 title={Be like a Goldfish, Don't Memorize! Mitigating Memorization in Generative LLMs},
3 author={Abhimanyu Hans and Yuxin Wen and Neel Jain and John Kirchenbauer and Hamid Kazemi and Prajwal Singhania and Siddharth Singh and Gowthami Somepalli and Jonas Geiping and Abhinav Bhatele and Tom Goldstein},
4 year={2024},
5 eprint={2406.10209},
6 archivePrefix={arXiv},
7}
