TylerHMC/GigaMIDI-starts-set
收藏资源简介:
Note on padding and training: This dataset contains some sequences that are already padded to a fixed length. If you train a causal language model and simply copy input_ids to labels without masking, the model will also learn to predict padding tokens, which can lead to artificially low loss values and misleading training results. To avoid this, you should mask padding tokens in the labels by setting them to -100 (so they are ignored by the loss function). For example: def causal_lm_collator(features): batch = base_collator(features) labels = batch["input_ids"].clone() labels[labels == tokenizer.pad_token_id] = -100 batch["labels"] = labels return batch This ensures that padding does not affect training and that reported loss values reflect real model performance. . --- dataset_info: features: - name: input_ids list: int32 length: 4096 - name: token_count dtype: int32 - name: source_file dtype: string - name: split dtype: string splits: - name: train num_bytes: 6810522632 num_examples: 415048 - name: validation num_bytes: 1705381772 num_examples: 103898 download_size: 8510172926 dataset_size: 8515904404 configs: - config_name: default data_files: - split: train path: data/train-* - split: validation path: data/validation-* ---




