utils.data.rl
utils.data.rl
Data handling specific to RL trainers.
Functions
| Name | Description |
|---|---|
| prepare_preference_datasets | Load and prepare preference datasets for RL training. |
prepare_preference_datasets
utils.data.rl.prepare_preference_datasets(cfg, tokenizer)Load and prepare preference datasets for RL training.
Loads training and evaluation datasets, handling preprocessing, caching, and deduplication as configured. Uses FileLock for distributed coordination.
Parameters
| Name | Type | Description | Default |
|---|---|---|---|
| cfg | DictDefault | Configuration object containing dataset and training settings. | required |
| tokenizer | PreTrainedTokenizer | Tokenizer to use for processing text. | required |
Returns
| Name | Type | Description |
|---|---|---|
| tuple[Dataset, Dataset | None] | Tuple of (train_dataset, eval_dataset). eval_dataset may be None if no evaluation dataset is configured. |