trl/pr_4949/en/data_utils.md

# Data Utilities

## prepare_multimodal_messages[[trl.prepare_multimodal_messages]]

#### trl.prepare_multimodal_messages[[trl.prepare_multimodal_messages]]

[Source](https://github.com/huggingface/trl/blob/vr_4949/trl/data_utils.py#L32)

Convert messages into a structured multimodal format and inject the provided images into the message contents.

Notes:
- When the input `messages` isn't already in the structured format, (i.e., all `"content"` values are strings),
  the function transforms them into the structured format by wrapping text in `{"type": "text", "text": ...}`
  and inserting `{"type": "image"}` placeholders for the images *before* the first user message.
- When the input `messages` is already in the structured format (i.e., all `"content"` values are lists of
  structured blocks), the function only fills in the actual images in the existing `{"type": "image"}`
  placeholders. If the number of placeholders does not match the number of provided images, an error is raised.

Example:
```python
# Input
[
    {"role": "user", "content": "What's in this image?"},
    {"role": "assistant", "content": "It looks like a cat."},
]

# Output, one image provided
[
    {"role": "user", "content": [{"type": "image", "image": }, {"type": "text", "text": "What's in this image?"}]},
    {"role": "assistant", "content": [{"type": "text", "text": "It looks like a cat."}]},
]
```

**Parameters:**

messages (`list[dict[str, Any]]`) : Messages with `"role"` and `"content"`. Content may be a raw string before transformation. List of messages a `"role"` key (`"system"`, `"user"`, or `"assistant"`) and a `"content"` key containing either a string or a list of structured blocks if already prepared.

images (`list`) : List of image objects to insert.

**Returns:**

``list[dict[str, Any]]``

A deep-copied list of messages where every `"content"` value is a list of structured
content blocks, and all `"image"` placeholders are populated with the corresponding image objects.

## prepare_multimodal_messages_vllm[[trl.prepare_multimodal_messages_vllm]]

#### trl.prepare_multimodal_messages_vllm[[trl.prepare_multimodal_messages_vllm]]

[Source](https://github.com/huggingface/trl/blob/vr_4949/trl/data_utils.py#L112)

Convert structured multimodal messages into a format compatible with vLLM. Replaces `"type": "image"` blocks with
`"type": "image_pil"` blocks, and `"image": Image` with `"image_pil": Image`.

Example:
```python
# Input
[{"role": "user", "content": [{"type": "image", "image": }, {"type": "text", "text": "What's in this image?"}]}]

# Output
[{"role": "user", "content": [{"type": "image_pil", "image_pil": }, {"type": "text", "text": "What's in this image?"}]}]
```

**Parameters:**

messages (`list[dict[str, Any]]`) : Messages with `"role"` and `"content"`. Content is expected to be a list of structured blocks.

**Returns:**

``list[dict[str, Any]]``

A deep-copied list of messages compatible with vLLM's expected input format.

## is_conversational[[trl.is_conversational]]

#### trl.is_conversational[[trl.is_conversational]]

[Source](https://github.com/huggingface/trl/blob/vr_4949/trl/data_utils.py#L145)

Check if the example is in a conversational format.

Examples:

```python
>>> example = {"prompt": [{"role": "user", "content": "What color is the sky?"}]}
>>> is_conversational(example)
True

>>> example = {"prompt": "The sky is"}
>>> is_conversational(example)
False
```

**Parameters:**

example (`dict[str, Any]`) : A single data entry of a dataset. The example can have different keys depending on the dataset type.

**Returns:**

``bool``

`True` if the data is in a conversational format, `False` otherwise.

## is_conversational_from_value[[trl.is_conversational_from_value]]

#### trl.is_conversational_from_value[[trl.is_conversational_from_value]]

[Source](https://github.com/huggingface/trl/blob/vr_4949/trl/data_utils.py#L897)

Check if the example is in a conversational format (from/value). Note that this format isn't recommended. Prefer
the ChatML format (role/content)

Examples:

```python
>>> example = {"conversations": [{"from": "user", "value": "What color is the sky?"}]}
>>> is_conversational_from_value(example)
True

>>> example = {"conversations": [{"role": "user", "content": "What color is the sky?"}]}
>>> is_conversational_from_value(example)
False

>>> example = {"conversations": "The sky is"}
>>> is_conversational_from_value(example)
False
```

**Parameters:**

example (`dict[str, Any]`) : A single data entry of a dataset. The example can have different keys depending on the dataset type.

**Returns:**

``bool``

`True` if the data is in a conversational Chatformat, `False` otherwise.

## apply_chat_template[[trl.apply_chat_template]]

#### trl.apply_chat_template[[trl.apply_chat_template]]

[Source](https://github.com/huggingface/trl/blob/vr_4949/trl/data_utils.py#L186)

Apply a chat template to a conversational example along with the schema for a list of functions in `tools`.

For more details, see [maybe_apply_chat_template()](/docs/trl/pr_4949/en/data_utils#trl.maybe_apply_chat_template).

## maybe_apply_chat_template[[trl.maybe_apply_chat_template]]

#### trl.maybe_apply_chat_template[[trl.maybe_apply_chat_template]]

[Source](https://github.com/huggingface/trl/blob/vr_4949/trl/data_utils.py#L319)

If the example is in a conversational format, apply a chat template to it.

Notes:
- This function does not alter the keys, except for Language modeling dataset, where `"messages"` is replaced
by `"text"`.

- In case of prompt-only data, if the last role is `"user"`, the generation prompt is added to the prompt.
Else, if the last role is `"assistant"`, the final message is continued.

Example:

```python
>>> from transformers import AutoTokenizer

>>> tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-3-mini-128k-instruct")
>>> example = {
...     "prompt": [{"role": "user", "content": "What color is the sky?"}],
...     "completion": [{"role": "assistant", "content": "It is blue."}],
... }
>>> apply_chat_template(example, tokenizer)
{'prompt': '\nWhat color is the sky?\n\n', 'completion': 'It is blue.\n'}
```

**Parameters:**

example (`dict[str, list[dict[str, str]]`) : Dictionary representing a single data entry of a conversational dataset. Each data entry can have different keys depending on the dataset type. The supported dataset types are:  - Language modeling dataset: `"messages"`. - Prompt-only dataset: `"prompt"`. - Prompt-completion dataset: `"prompt"` and `"completion"`. - Preference dataset: `"prompt"`, `"chosen"`, and `"rejected"`. - Preference dataset with implicit prompt: `"chosen"` and `"rejected"`. - Unpaired preference dataset: `"prompt"`, `"completion"`, and `"label"`.  For keys `"messages"`, `"prompt"`, `"chosen"`, `"rejected"`, and `"completion"`, the values are lists of messages, where each message is a dictionary with keys `"role"` and `"content"`. Additionally, the example may contain a `"chat_template_kwargs"` key, which is a dictionary of additional keyword arguments to pass to the chat template renderer.

tokenizer ([PreTrainedTokenizerBase](https://huggingface.co/docs/transformers/main/en/internal/tokenization_utils#transformers.PreTrainedTokenizerBase)) : Tokenizer to apply the chat template with.

tools (`list[dict | Callable]`, *optional*) : A list of tools (callable functions) that will be accessible to the model. If the template does not support function calling, this argument will have no effect.

- ****template_kwargs** (`Any`, *optional*) : Additional kwargs to pass to the template renderer. Will be accessible by the chat template.

**Returns:**

``dict[str, str]``

Formatted example with the chat template applied.

## maybe_convert_to_chatml[[trl.maybe_convert_to_chatml]]

#### trl.maybe_convert_to_chatml[[trl.maybe_convert_to_chatml]]

[Source](https://github.com/huggingface/trl/blob/vr_4949/trl/data_utils.py#L937)

Convert a conversational dataset with fields `from` and `value` to ChatML format.

This function modifies conversational data to align with OpenAI's ChatML format:
- Replaces the key `"from"` with `"role"` in message dictionaries.
- Replaces the key `"value"` with `"content"` in message dictionaries.
- Renames `"conversations"` to `"messages"` for consistency with ChatML.

Example:
```python
>>> from trl import maybe_convert_to_chatml

>>> example = {
...     "conversations": [
...         {"from": "user", "value": "What color is the sky?"},
...         {"from": "assistant", "value": "It is blue."},
...     ]
... }
>>> maybe_convert_to_chatml(example)
{'messages': [{'role': 'user', 'content': 'What color is the sky?'},
              {'role': 'assistant', 'content': 'It is blue.'}]}
```

**Parameters:**

example (`dict[str, list]`) : A single data entry containing a list of messages.

**Returns:**

``dict[str, list]``

Example reformatted to ChatML style.

## extract_prompt[[trl.extract_prompt]]

#### trl.extract_prompt[[trl.extract_prompt]]

[Source](https://github.com/huggingface/trl/blob/vr_4949/trl/data_utils.py#L488)

Extracts the shared prompt from a preference data example, where the prompt is implicit within both the chosen and
rejected completions.

For more details, see [maybe_extract_prompt()](/docs/trl/pr_4949/en/data_utils#trl.maybe_extract_prompt).

## maybe_extract_prompt[[trl.maybe_extract_prompt]]

#### trl.maybe_extract_prompt[[trl.maybe_extract_prompt]]

[Source](https://github.com/huggingface/trl/blob/vr_4949/trl/data_utils.py#L507)

Extracts the shared prompt from a preference data example, where the prompt is implicit within both the chosen and
rejected completions.

If the example already contains a `"prompt"` key, the function returns the example as is. Else, the function
identifies the longest common sequence (prefix) of conversation turns between the "chosen" and "rejected"
completions and extracts this as the prompt. It then removes this prompt from the respective "chosen" and
"rejected" completions.

Examples:

```python
>>> example = {
...     "chosen": [
...         {"role": "user", "content": "What color is the sky?"},
...         {"role": "assistant", "content": "It is blue."},
...     ],
...     "rejected": [
...         {"role": "user", "content": "What color is the sky?"},
...         {"role": "assistant", "content": "It is green."},
...     ],
... }
>>> extract_prompt(example)
{'prompt': [{'role': 'user', 'content': 'What color is the sky?'}],
 'chosen': [{'role': 'assistant', 'content': 'It is blue.'}],
 'rejected': [{'role': 'assistant', 'content': 'It is green.'}]}
```

Or, with the `map` method of [Dataset](https://huggingface.co/docs/datasets/main/en/package_reference/main_classes#datasets.Dataset):

```python
>>> from trl import extract_prompt
>>> from datasets import Dataset

>>> dataset_dict = {
...     "chosen": [
...         [
...             {"role": "user", "content": "What color is the sky?"},
...             {"role": "assistant", "content": "It is blue."},
...         ],
...         [
...             {"role": "user", "content": "Where is the sun?"},
...             {"role": "assistant", "content": "In the sky."},
...         ],
...     ],
...     "rejected": [
...         [
...             {"role": "user", "content": "What color is the sky?"},
...             {"role": "assistant", "content": "It is green."},
...         ],
...         [
...             {"role": "user", "content": "Where is the sun?"},
...             {"role": "assistant", "content": "In the sea."},
...         ],
...     ],
... }
>>> dataset = Dataset.from_dict(dataset_dict)
>>> dataset = dataset.map(extract_prompt)
>>> dataset[0]
{'prompt': [{'role': 'user', 'content': 'What color is the sky?'}],
 'chosen': [{'role': 'assistant', 'content': 'It is blue.'}],
 'rejected': [{'role': 'assistant', 'content': 'It is green.'}]}
```

**Parameters:**

example (`dict[str, list]`) : A dictionary representing a single data entry in the preference dataset. It must contain the keys `"chosen"` and `"rejected"`, where each value is either conversational or standard (`str`).

**Returns:**

``dict[str, list]``

A dictionary containing:
- `"prompt"`: The longest common prefix between the "chosen" and "rejected" completions.
- `"chosen"`: The remainder of the "chosen" completion, with the prompt removed.
- `"rejected"`: The remainder of the "rejected" completion, with the prompt removed.

## unpair_preference_dataset[[trl.unpair_preference_dataset]]

#### trl.unpair_preference_dataset[[trl.unpair_preference_dataset]]

[Source](https://github.com/huggingface/trl/blob/vr_4949/trl/data_utils.py#L394)

Unpair a preference dataset.

Example:

```python
>>> from datasets import Dataset

>>> dataset_dict = {
...     "prompt": ["The sky is", "The sun is"],
...     "chosen": [" blue.", "in the sky."],
...     "rejected": [" green.", " in the sea."],
... }
>>> dataset = Dataset.from_dict(dataset_dict)
>>> dataset = unpair_preference_dataset(dataset)
>>> dataset
Dataset({
    features: ['prompt', 'completion', 'label'],
    num_rows: 4
})

>>> dataset[0]
{'prompt': 'The sky is', 'completion': ' blue.', 'label': True}
```

**Parameters:**

dataset ([Dataset](https://huggingface.co/docs/datasets/main/en/package_reference/main_classes#datasets.Dataset) or [DatasetDict](https://huggingface.co/docs/datasets/main/en/package_reference/main_classes#datasets.DatasetDict)) : Preference dataset to unpair. The dataset must have columns `"chosen"`, `"rejected"` and optionally `"prompt"`.

num_proc (`int`, *optional*) : Number of processes to use for processing the dataset.

desc (`str`, *optional*) : Meaningful description to be displayed alongside with the progress bar while mapping examples.

**Returns:**

`[Dataset](https://huggingface.co/docs/datasets/main/en/package_reference/main_classes#datasets.Dataset)`

The unpaired preference dataset.

## maybe_unpair_preference_dataset[[trl.maybe_unpair_preference_dataset]]

#### trl.maybe_unpair_preference_dataset[[trl.maybe_unpair_preference_dataset]]

[Source](https://github.com/huggingface/trl/blob/vr_4949/trl/data_utils.py#L437)

Unpair a preference dataset if it is paired.

Example:

```python
>>> from datasets import Dataset

>>> dataset_dict = {
...     "prompt": ["The sky is", "The sun is"],
...     "chosen": [" blue.", "in the sky."],
...     "rejected": [" green.", " in the sea."],
... }
>>> dataset = Dataset.from_dict(dataset_dict)
>>> dataset = unpair_preference_dataset(dataset)
>>> dataset
Dataset({
    features: ['prompt', 'completion', 'label'],
    num_rows: 4
})

>>> dataset[0]
{'prompt': 'The sky is', 'completion': ' blue.', 'label': True}
```

**Parameters:**

dataset ([Dataset](https://huggingface.co/docs/datasets/main/en/package_reference/main_classes#datasets.Dataset) or [DatasetDict](https://huggingface.co/docs/datasets/main/en/package_reference/main_classes#datasets.DatasetDict)) : Preference dataset to unpair. The dataset must have columns `"chosen"`, `"rejected"` and optionally `"prompt"`.

num_proc (`int`, *optional*) : Number of processes to use for processing the dataset.

desc (`str`, *optional*) : Meaningful description to be displayed alongside with the progress bar while mapping examples.

**Returns:**

`[Dataset](https://huggingface.co/docs/datasets/main/en/package_reference/main_classes#datasets.Dataset) or [DatasetDict](https://huggingface.co/docs/datasets/main/en/package_reference/main_classes#datasets.DatasetDict)`

The unpaired preference dataset if it was paired, otherwise
the original dataset.

## pack_dataset[[trl.pack_dataset]]

#### trl.pack_dataset[[trl.pack_dataset]]

[Source](https://github.com/huggingface/trl/blob/vr_4949/trl/data_utils.py#L778)

Pack sequences in a dataset into chunks of size `seq_length`.

Example:
```python
>>> from datasets import Dataset
>>> from trl import pack_dataset

>>> examples = {
...     "input_ids": [[1, 2, 3, 4, 5], [6, 7], [8, 9, 10], [11]],
...     "attention_mask": [[1, 1, 1, 0, 0], [1, 0], [1, 1, 0], [1]],
... }
>>> dataset = Dataset.from_dict(examples)
>>> packed_dataset = pack_dataset(dataset, seq_length=4, strategy="bfd")
>>> packed_dataset[:]
{'input_ids': [[1, 2, 3, 4], [8, 9, 10, 5], [6, 7, 11]],
 'attention_mask': [[1, 1, 1, 0], [0, 1, 1, 0], [1, 1, 1]],
 'seq_lengths': [[4], [3, 1], [2, 1]]}
```

**Parameters:**

dataset ([Dataset](https://huggingface.co/docs/datasets/main/en/package_reference/main_classes#datasets.Dataset) or [DatasetDict](https://huggingface.co/docs/datasets/main/en/package_reference/main_classes#datasets.DatasetDict)) : Dataset to pack

seq_length (`int`) : Target sequence length to pack to.

strategy (`str`, *optional*, defaults to `"bfd"`) : Packing strategy to use. Can be either:  - `"bfd"` (Best Fit Decreasing): Slower but preserves sequence boundaries inside each packed sample. If a single sequence exceeds `seq_length` it is split into multiple samples. - `"wrapped"`: Faster but more aggressive. Ignores sequence boundaries and will cut sequences in the middle to completely fill each packed sequence with data.

map_kwargs (`dict`, *optional*) : Additional keyword arguments to pass to the dataset's map method when packing examples.

**Returns:**

`[Dataset](https://huggingface.co/docs/datasets/main/en/package_reference/main_classes#datasets.Dataset) or [DatasetDict](https://huggingface.co/docs/datasets/main/en/package_reference/main_classes#datasets.DatasetDict)`

The dataset with packed sequences. The number of examples
may decrease as sequences are combined.

## truncate_dataset[[trl.truncate_dataset]]

#### trl.truncate_dataset[[trl.truncate_dataset]]

[Source](https://github.com/huggingface/trl/blob/vr_4949/trl/data_utils.py#L834)

Truncate sequences in a dataset to a specified `max_length`.

Example:
```python
>>> from datasets import Dataset

>>> examples = {
...     "input_ids": [[1, 2, 3], [4, 5, 6, 7], [8]],
...     "attention_mask": [[0, 1, 1], [0, 0, 1, 1], [1]],
... }
>>> dataset = Dataset.from_dict(examples)
>>> truncated_dataset = truncate_dataset(dataset, max_length=2)
>>> truncated_dataset[:]
{'input_ids': [[1, 2], [4, 5], [8]],
 'attention_mask': [[0, 1], [0, 0], [1]]}
```

**Parameters:**

dataset ([Dataset](https://huggingface.co/docs/datasets/main/en/package_reference/main_classes#datasets.Dataset) or [DatasetDict](https://huggingface.co/docs/datasets/main/en/package_reference/main_classes#datasets.DatasetDict)) : Dataset to truncate.

max_length (`int`) : Maximum sequence length to truncate to.

map_kwargs (`dict`, *optional*) : Additional keyword arguments to pass to the dataset's map method when truncating examples.

**Returns:**

`[Dataset](https://huggingface.co/docs/datasets/main/en/package_reference/main_classes#datasets.Dataset) or [DatasetDict](https://huggingface.co/docs/datasets/main/en/package_reference/main_classes#datasets.DatasetDict)`

The dataset with truncated sequences.