KD-gpt2-120M

paper | code

KD-gpt2-120M is a gpt2-base (120M) model distilled from gpt2-xlarge (1.5B) on databricks-dolly-15k with token-level forward KLD.

It is used as a baseline for MiniLLM.

Other Baselines

• SFT w/o KD
• SeqKD

Citation

@inproceedings{minillm,
  title={MiniLLM: Knowledge Distillation of Large Language Models},
  author={Gu, Yuxian and Dong, Li and Wei, Furu and Huang, Minlie},
  booktitle={Proceedings of ICLR},
  year={2024}
}