torchtune.data¶

Text templates¶

Templates for instruct prompts and chat prompts. Includes some specific formatting for difference datasets and models.

`GrammarErrorCorrectionTemplate`	A prompt template for grammar error correction tasks.
`SummarizeTemplate`	A prompt template for summarization tasks.
`QuestionAnswerTemplate`	A prompt template for question answering tasks.
`PromptTemplate`	Quickly define a custom prompt template by passing in a dictionary mapping role to the prepend and append tags. For example, to achieve the following prompt template::.
`PromptTemplateInterface`	Interface for prompt templates.
`ChatMLTemplate`	OpenAI's Chat Markup Language used by their chat models.

`Message`	This class represents individual messages in a fine-tuning dataset.
`Role`	alias of `Literal`['system', 'user', 'assistant', 'ipython']

Converts data from common schema and conversation JSON formats into a list of torchtune Message.

`InputOutputToMessages`	Message transform class that converts a single sample with "input" and "output" fields, (or equivalent fields specified in column_map) to user and assistant messages, respectively. This is useful for datasets that have two columns, one containing the user prompt string and the other containing the model response string::.
`ShareGPTToMessages`	Convert a single chat sample adhering to the ShareGPT JSON structure to torchtune's `Message` structure.
`OpenAIToMessages`	Convert a single chat sample adhering to the OpenAI chat completion JSON structure to torchtune's `Message` structure.
`ChosenRejectedToMessages`	Transform for converting a single sample from datasets with "chosen" and "rejected" columns containing conversations to a list of chosen and rejected messages. For example::.
`AlpacaToMessages`	Message transform class for Alpaca-style datasets with "instruction", "input", and "output" (or equivalent fields specified in column_map) columns.

Collaters used to collect samples into batches and handle any padding.

`padded_collate`	A generic padding collation function which pads `keys_to_pad` entries in a batch of sequences from the given `pad_direction` to the maximum sequence length for each entry in the batch.
`padded_collate_tiled_images_and_mask`	Pad a batch of text sequences, tiled image tensors, aspect ratios, and cross attention masks.
`padded_collate_sft`	Pad a batch of sequences to the longest sequence length in the batch, and convert integer lists to tensors.
`padded_collate_dpo`	Pad a batch of sequences for Direct Preference Optimization (DPO).
`left_pad_sequence`	This function is identical to `torch.nn.utils.rnn.pad_sequence()`, but instead pads a list of variable length Tensors from the left to the length of the longest sequence.

Miscellaneous helper functions used in modifying data.

`validate_messages`	Given a list of messages, ensure that messages form a valid back-and-forth conversation.
`truncate`	Truncate a list of tokens to a maximum length.
`load_image`	Convenience method to load an image in PIL format from a local file path or remote source.
`format_content_with_images`	Given a raw text string, split by the specified `image_tag` and form into list of dictionaries to be used in the `Message` content field.