Release v0.5.1

Updates:

• Fixed minor SimCSE modeling bugs
• Rename layer names in Question Answering models

Release v0.5.0

Updates:

• Introduce smile-datasets to build data pipeline

Release v0.4.9

Updates:

• Rename Examples and Datasets for consistency
• Rename Callbacks for consistency
• Refactoring Adapters for better subclassing
• Add BertAdapterForLangboatMengzi for load pretrained weights from Langboat/Mengzi

Release v0.4.8

Updates:

• Fixed building dataset from CoNLL files for token classification tasks
• Rename method _parse_jsonl to _parse_instances_to_examples for dataset builders
• Fixed minor arguments error

Release v0.4.7

Changes:

• Fixed confusing warnings in adapters
• Add BertForMaskedLanguageModel and it's dataset builder MaskedLanguageModelDataset

Release v0.4.6

Changes:

• Add Distiller for model distillation
• Add support for difference padding strategies:
  • bucket: group sequence by length, and then padding sequence by max sequence length of the same batch
  • batch: padding sequence by max sequence length of this batch
  • fixed: padding sequence to a fixed max sequence length over all examples

Release 0.4.5

Updates:

• Simplify dataset builders
• Add Sentiment Analysis models & dataset builder

Release v0.4.4

Updates:

• Refactoring datasets for different tasks, add new from_jsonl_files implementations
• Add metric callbacks for different tasks, including SpearmanForSentenceEmbedding, ExactMatchForQuestionAnswering, F1ForQuestionAnswering,SeqEvalForTokenClassification
• Fixed some error in datasets

Release v0.4.3

Updates:

• Remove from_jsonl_files method from dataset builders for simplicity and consistency

Release v0.4.2

Updates:

• Add dataset builders for question answering tasks
• Add dataset builders for simcse model
• Add dataset builders for sequence classification tasks
• Add dataset builders for token classification tasks