MCPcopy Create free account
hub / github.com/THUDM/GLM / wrap_dataset

Function wrap_dataset

data_utils/__init__.py:184–198  ·  view source on GitHub ↗
(dataset)

Source from the content-addressed store, hash-verified

182
183 # Split dataset into train/val/test (and wrap bert dataset)
184 def wrap_dataset(dataset):
185 if ds_type.lower() == 'bert':
186 presplit_sentences = kwargs['presplit_sentences'] if 'presplit_sentences' in kwargs else False
187 dataset = BertSentencepairDataset(dataset, max_seq_len=seq_length, presplit_sentences=presplit_sentences)
188 elif ds_type.lower() == 'gpt-xl':
189 assert pre_tokenize
190 dataset = XLDataset(dataset, tokenizer, max_seq_len=seq_length, mem_len=mem_length,
191 sample_across_doc=not sample_one_document)
192 elif ds_type.lower() == 'gpt2':
193 dataset = GPT2Dataset(dataset, tokenizer, max_seq_len=seq_length, sample_across_doc=not sample_one_document)
194 elif ds_type.lower() == 'block':
195 dataset = BlockDataset(dataset, tokenizer, max_seq_len=seq_length,
196 sample_across_doc=not sample_one_document, filter_english=filter_english,
197 non_sentence_start=non_sentence_start)
198 return dataset
199
200 if should_split(split):
201 ds = split_ds(ds, split, shuffle=shuffle, save_splits=save_splits, load_splits=load_splits)

Callers 1

make_datasetFunction · 0.85

Calls 4

XLDatasetClass · 0.85
GPT2DatasetClass · 0.85
BlockDatasetClass · 0.85

Tested by

no test coverage detected