Shortcuts

preference_dataset

torchtune.datasets.preference_dataset(tokenizer: ModelTokenizer, *, source: str, column_map: Optional[Dict[str, str]] = None, train_on_input: bool = False, new_system_prompt: Optional[str] = None, split: str = 'train', **load_dataset_kwargs: Dict[str, Any]) PreferenceDataset[source]

Configures a custom preference dataset comprising interactions between user and model assistant.

This builder function can be used to configure a custom preference dataset directly from the yaml config as an alternative to PreferenceDataset, as it is made to be config friendly.

This function requires the dataset to have “chosen” and “rejected” columns. A single sample will share an identical system +/ user prompt between both “chosen” and “rejected” columns, followed by one or multiple turns of user and assistant messages:

|  chosen                                |  rejected                              |
|----------------------------------------|----------------------------------------|
| [{"role": "user", "content": Q1},      | [{"role": "user", "content": Q1},      |
|  {"role": "assistant", "content": C1}] |  {"role": "assistant", "content": R1}] |

This example will be converted to:

chosen_messages = [
    Message(role="user", content="Q1"),
    Message(role="assistant", content="C1"),
]

rejected_messages = [
    Message(role="user", content="Q1"),
    Message(role="assistant", content="R1"),
]

These lists of messages are then tokenized for model training. Currently, this function only supports conversations identical to OpenAIToMessages, and does not support custom message formats.

If your dataset does not follow this format, we recommend creating a custom message transform similar to ChosenRejectedToMessages and using it in a custom dataset builder function similar to preference_dataset.

Masking of the prompt during training is controlled by the train_on_input flag, which is: set to False by default.

  • If train_on_input is True, the prompt is used during training and contributes to the loss.

  • If train_on_input is False, the prompt is masked out (tokens replaced with -100).

Parameters:
  • tokenizer (ModelTokenizer) – Tokenizer used by the model that implements the tokenize_messages method.

  • source (str) – path to dataset repository on Hugging Face. For local datasets, define source as the data file type (e.g. “json”, “csv”, “text”), pass in the filepath in data_files, and set split="train". See Hugging Face’s load_dataset for more details.

  • column_map (Optional[Dict[str, str]]) – a mapping from the expected columns “chosen” and “rejected” in the message transform ChosenRejectedToMessages to the new column names in the dataset. Keys should be “chosen” and “rejected” and values should be the actual column names. If None, keep the default columns “chosen” and “rejected”.

  • train_on_input (bool) – Whether the model is trained on the prompt or not. Default is False.

  • new_system_prompt (Optional[str]) – if specified, prepend a system message to every sample for both chosen and rejected. This can serve as instructions to guide the model response. Setting this will OVERRIDE any system messages already present in the dataset. Default is None.

  • split (str) – split argument for datasets.load_dataset. You can use this argument to load a subset of a given split, e.g. split="train[:10%]". Default is “train”.

  • **load_dataset_kwargs (Dict[str, Any]) – additional keyword arguments to pass to load_dataset.

Examples:

my_preference_dataset.json
[
    {
        "chosen_conversations": [
            {
                "content": "What do I do when I have a hole in my trousers?",
                "role": "user"
            },
            { "content": "Fix the hole.", "role": "assistant" }
        ],
        "rejected_conversations": [
            {
                "content": "What do I do when I have a hole in my trousers?",
                "role": "user"
            },
            { "content": "Take them off.", "role": "assistant" }
        ]
    }
]
>>> from torchtune.datasets import preference_dataset
>>> column_map = {
...     "chosen": "chosen_conversations",
...     "rejected": "rejected_conversations"
>>> }
>>> dataset = preference_dataset(
...     tokenizer=tokenizer,
...     source="json",
...     column_map=column_map,
...     data_files="my_preference_dataset.json",
...     train_on_input=False,
...     split="train",
>>> )
>>> tokenizer.decode(dataset[0]["chosen_input_ids"], skip_special_tokens=True)
What do I do when I have a hole in my trousers?Fix the hole.
>>> tokenizer.decode(dataset[0]["rejected_input_ids"], skip_special_tokens=True)
What do I do when I have a hole in my trousers?Take them off.

This can also be accomplished via the yaml config:

dataset:
  _component_: torchtune.datasets.preference_dataset
  source: json
  data_files: my_preference_dataset.json
  column_map:
    chosen: chosen_conversations
    rejected: rejected_conversations
  train_on_input: False
  split: train
Returns:

The preference dataset built from source paired data.

Return type:

PreferenceDataset

Docs

Access comprehensive developer documentation for PyTorch

View Docs

Tutorials

Get in-depth tutorials for beginners and advanced developers

View Tutorials

Resources

Find development resources and get your questions answered

View Resources