chat_dataset

torchtune.datasets.chat_dataset(tokenizer: ModelTokenizer, *, source: str, conversation_column: str, conversation_style: str, train_on_input: bool = False, packed: bool = False, **load_dataset_kwargs: Dict[str, Any]) → Union[SFTDataset, PackedDataset]

Configure a custom dataset with conversations between user and model assistant.

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

The dataset is expected to contain a single column with the conversations:

|  conversations                         |
|----------------------------------------|
| [{"role": "user", "content": Q1},      |
|  {"role": "assistant", "content": A1}] |

This will be converted to:

messages = [
    Message(role="user", content="Q1"),
    Message(role="assistant", content="A1"),
]

This list of messages is then tokenized for model training.

You may have a different structure for your conversations, such as different role names or different keys in the json structure. You can use the conversation_style parameter to choose from standard formats such as “sharegpt” (see ShareGPTToMessages) or “json” (see JSONToMessages). If your dataset is not in one of these formats, we recommend creating a custom message transform and using it in a custom dataset builder function similar to chat_dataset.

If your column names are different, use the conversation_column parameter to point towards the column with the conversations.

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.

• conversation_column (str) – name of column containing the conversations.

• conversation_style (str) – string specifying expected style of conversations in the dataset for automatic conversion to the Message structure. Supported styles are: “sharegpt”, “json”

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

• packed (bool) – Whether or not to pack the dataset to max_seq_len prior to training. Default is False.

• **load_dataset_kwargs (Dict[str, Any]) – additional keyword arguments to pass to load_dataset, such as data_files or split.

Examples:

my_dataset.json
[
    {
        "conversations": [
            {
                "from": "human",
                "value": "What time is it in London?",
            },
            {
                "from": "gpt",
                "value": "It is 10:00 AM in London.",
            },
        ],
    },
    {
        "conversations": [
            ...
        ],
    },
    ...,
]

>>> from torchtune.datasets import chat_dataset
>>> dataset = chat_dataset(
...     tokenizer=tokenizer,
...     source="json",
...     data_files="my_dataset.json",
...     conversation_column="conversations",
...     conversation_style="sharegpt",
...     train_on_input=False,
...     packed=False,
...     split="train",
... )
>>> tokens = dataset[0]["tokens"]
>>> tokenizer.decode(tokens)
"What time is it in London?It is 10:00 AM in London."

This can also be accomplished via the yaml config:

dataset:
  _component_: torchtune.datasets.chat_dataset
  source: json
  data_files: my_dataset.json
  conversation_column: conversations
  conversation_style: sharegpt
  train_on_input: False
  packed: False
  split: train

Returns:

the configured SFTDataset

or PackedDataset if packed=True

Return type:

Union[SFTDataset, PackedDataset]

Raises:

ValueError – if the conversation format is not supported