Dataset V4.0

[nicolay-r]

• Uniform modeling across all utterances from all books of the same set  #28
• Use Masked _ character names instead of their IDS. #18
• Limitation on HLA utilized for selection; auhtors consider 40 highest HLA per speaker, among which we perform random selection #27
  
• Authors use non-intersected speaker sets! #26
• Switch to 5 folding parts instead of 10
  
• Speaker set is expected to be smaller #25
  
• Why we have a few amount of dialogs per speaker than expected 100? (v. 3.1) #29
• The amount of training samples reprent ~24% of the same amount in ALOHA studies #30
• Filtering response utterances that at least mention $k$ speakers #35
• Fixed to Persona: none in the case when persona is absent (df6bfd0)
• Dialogue lines were treated with meta-information #36
• oversampled candidates are the same [bug] (04b18fa)
• fixed bug with absense of the newline after a set of oversampled examples (c8f400b)
• spectrums are not randomly selected (41addd9)