BaohaoLiao / NLP-reproduction

Offer straightforward guidance to reproduce the results of NLP papers.

NLP-reproduction

This repository aims to offer straightforward guidance to reproduce the results from NLP papers. We only focus on the tasks that are not easy to reproduce with the official guidance, and try to obtain almost the same scores reported in papers.

Your ⭐ motivate us to continue!!!

Summarization

• XSum

  Method	R1	R2	RL
  BART paper	45.14	22.27	37.25
  evaluate fine-tuned BART (our)	45.20	21.91	36.69
  finetune pre-tuned BART (our)	44.91	21.91	36.73

  • Download and split [2023.03.09]
  • Evaluate fine-tuned BART on fairseq [2023.03.09]
  • Finetune pre-trained BART on fairseq [2023.03.13]

• CNN/DM

  Method	R1	R2	RL
  BART paper	44.16	21.28	40.90
  evaluate fine-tuned BART (our)	44.26	21.31	41.06
  finetune pre-tuned BART (our)	44.76	21.76	41.49

  • Download and preprocess [2023.03.10]
  • Evaluate fine-tuned BART on fairseq [2023.03.10]
  • Finetune pre-trained BART on fairseq [2023.03.11]

Translation

• WMT16 En-Ro
  • TODO:
    • Data preprocessing
    • Fintune mBART-CC25
    • post-processing script