princeton-nlp/DinkyTrain

Issues

• Perplexity numbers for masking rates reported in Table 3.

  #10 opened 7 months ago by raotnameh
  1

• While install dependencies ERROR: Could not find a version that satisfies the requirement hydra-core<1.1,>=1.0.7 (from fairseq) (from versions: none) ERROR: No matching distribution found for hydra-core<1.1,>=1.0.7

  #9 opened 2 years ago by henrywang0314
  1

• Could you please tell how to set the hyparameters of the GLUE?

  #8 opened 2 years ago by leoozy
  2

• Roberta recipe in your paper is different from the original recipe

  #4 opened 2 years ago by BaohaoLiao
  3

• fairseq-train: error: argument --arch/-a: invalid choice: 'deepspeed_roberta_large'

  #7 opened 2 years ago by leoozy
  2

• Converting fairseq models to huggingface fails for models trained using DeepSpeed

  #5 opened 3 years ago by carlosejimenez
  0

• Why do you use last checkpoint for validation rather than best checkpoint?

  #3 opened 3 years ago by BaohaoLiao
  4

• Prediction values of the STS-B test set are not in 0~5

  #2 opened 3 years ago by BaohaoLiao
  4

• Why do you use both layer_norm for embedding and pre-norm at the same time?

  #1 opened 3 years ago by BaohaoLiao
  2