pre-tokenize the dataset before training; only collate during training
(dataset, tokenizer)
| 103 | dataset_text_field = "prompt" |
| 104 | |
| 105 | def prepare_dataset(dataset, tokenizer): |
| 106 | """pre-tokenize the dataset before training; only collate during training""" |
| 107 | |
| 108 | def tokenize(element): |
| 109 | outputs = tokenizer( |
| 110 | element[dataset_text_field], |
| 111 | padding=False, |
| 112 | ) |
| 113 | return {"input_ids": outputs["input_ids"]} |
| 114 | |
| 115 | return dataset.map( |
| 116 | tokenize, |
| 117 | batched=True, |
| 118 | remove_columns=dataset.column_names, |
| 119 | num_proc=training_args.dataset_num_proc, |
| 120 | ) |
| 121 | |
| 122 | # Compute that only on the main process for faster data processing. |
| 123 | # see: https://github.com/huggingface/trl/pull/1255 |