trl/pr_4305/en/data_utils.md

# Data Utilities

## prepare_multimodal_messages[[trl.prepare_multimodal_messages]]

<div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8">


<docstring><name>trl.prepare_multimodal_messages</name><anchor>trl.prepare_multimodal_messages</anchor><source>https://github.com/huggingface/trl/blob/vr_4305/trl/data_utils.py#L31</source><parameters>[{"name": "messages", "val": ": list"}, {"name": "num_images", "val": ": int"}]</parameters><paramsdesc>- **messages** (`list[dict[str, Any]]`) --
  Messages with `"role"` and `"content"`. Content may be a raw string before transformation.
- **num_images** (`int`) --
  Number of images to include in the first user message. This is used to determine how many image
  placeholders to add.</paramsdesc><paramgroups>0</paramgroups></docstring>

Convert messages into a structured multimodal format if needed.

Each message's content is transformed from a raw string into a list of typed parts. The first user message is
prefixed with an image placeholder, while all other user and assistant messages are wrapped as text entries.



<ExampleCodeBlock anchor="trl.prepare_multimodal_messages.example">

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

# Output (num_images=1)
[
    {"role": "user", "content": [{"type": "image"}, {"type": "text", "text": "What's in this image?"}]},
    {"role": "assistant", "content": [{"type": "text", "text": "It looks like a cat."}]},
]
```

</ExampleCodeBlock>


</div>

## is_conversational[[trl.is_conversational]]

<div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8">


<docstring><name>trl.is_conversational</name><anchor>trl.is_conversational</anchor><source>https://github.com/huggingface/trl/blob/vr_4305/trl/data_utils.py#L79</source><parameters>[{"name": "example", "val": ": dict"}]</parameters><paramsdesc>- **example** (`dict[str, Any]`) --
  A single data entry of a dataset. The example can have different keys depending on the dataset type.</paramsdesc><paramgroups>0</paramgroups><rettype>`bool`</rettype><retdesc>`True` if the data is in a conversational format, `False` otherwise.</retdesc></docstring>

Check if the example is in a conversational format.







<ExampleCodeBlock anchor="trl.is_conversational.example">

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
```

</ExampleCodeBlock>


</div>

## is_conversational_from_value[[trl.is_conversational_from_value]]

<div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8">


<docstring><name>trl.is_conversational_from_value</name><anchor>trl.is_conversational_from_value</anchor><source>https://github.com/huggingface/trl/blob/vr_4305/trl/data_utils.py#L782</source><parameters>[{"name": "example", "val": ": dict"}]</parameters><paramsdesc>- **example** (`dict[str, Any]`) --
  A single data entry of a dataset. The example can have different keys depending on the dataset type.</paramsdesc><paramgroups>0</paramgroups><rettype>`bool`</rettype><retdesc>`True` if the data is in a conversational Chatformat, `False` otherwise.</retdesc></docstring>

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







<ExampleCodeBlock anchor="trl.is_conversational_from_value.example">

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
```

</ExampleCodeBlock>


</div>

## apply_chat_template[[trl.apply_chat_template]]

<div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8">


<docstring><name>trl.apply_chat_template</name><anchor>trl.apply_chat_template</anchor><source>https://github.com/huggingface/trl/blob/vr_4305/trl/data_utils.py#L120</source><parameters>[{"name": "example", "val": ": dict"}, {"name": "tokenizer", "val": ": typing.Union[transformers.tokenization_utils_base.PreTrainedTokenizerBase, transformers.processing_utils.ProcessorMixin]"}, {"name": "tools", "val": ": typing.Optional[list[typing.Union[dict, typing.Callable]]] = None"}, {"name": "**template_kwargs", "val": ""}]</parameters></docstring>

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_4305/en/data_utils#trl.maybe_apply_chat_template).


</div>

## maybe_apply_chat_template[[trl.maybe_apply_chat_template]]

<div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8">


<docstring><name>trl.maybe_apply_chat_template</name><anchor>trl.maybe_apply_chat_template</anchor><source>https://github.com/huggingface/trl/blob/vr_4305/trl/data_utils.py#L249</source><parameters>[{"name": "example", "val": ": dict"}, {"name": "tokenizer", "val": ": PreTrainedTokenizerBase"}, {"name": "tools", "val": ": typing.Optional[list[typing.Union[dict, typing.Callable]]] = None"}, {"name": "**template_kwargs", "val": ": typing.Any"}]</parameters><paramsdesc>- **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[Union[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.</paramsdesc><paramgroups>0</paramgroups><rettype>`dict[str, str]`</rettype><retdesc>Formatted example with the chat template applied.</retdesc></docstring>

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.

<ExampleCodeBlock anchor="trl.maybe_apply_chat_template.example">

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': '<|user|>\nWhat color is the sky?<|end|>\n<|assistant|>\n', 'completion': 'It is blue.<|end|>\n'}
```

</ExampleCodeBlock>


</div>

## maybe_convert_to_chatml[[trl.maybe_convert_to_chatml]]

<div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8">


<docstring><name>trl.maybe_convert_to_chatml</name><anchor>trl.maybe_convert_to_chatml</anchor><source>https://github.com/huggingface/trl/blob/vr_4305/trl/data_utils.py#L822</source><parameters>[{"name": "example", "val": ": dict"}]</parameters><paramsdesc>- **example** (`dict[str, list]`) --
  A single data entry containing a list of messages.</paramsdesc><paramgroups>0</paramgroups><rettype>`dict[str, list]`</rettype><retdesc>Example reformatted to ChatML style.</retdesc></docstring>

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.







<ExampleCodeBlock anchor="trl.maybe_convert_to_chatml.example">

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.'}]}
```

</ExampleCodeBlock>


</div>

## extract_prompt[[trl.extract_prompt]]

<div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8">


<docstring><name>trl.extract_prompt</name><anchor>trl.extract_prompt</anchor><source>https://github.com/huggingface/trl/blob/vr_4305/trl/data_utils.py#L418</source><parameters>[{"name": "example", "val": ": dict"}]</parameters></docstring>

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_4305/en/data_utils#trl.maybe_extract_prompt).


</div>

## maybe_extract_prompt[[trl.maybe_extract_prompt]]

<div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8">


<docstring><name>trl.maybe_extract_prompt</name><anchor>trl.maybe_extract_prompt</anchor><source>https://github.com/huggingface/trl/blob/vr_4305/trl/data_utils.py#L437</source><parameters>[{"name": "example", "val": ": dict"}]</parameters><paramsdesc>- **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`).</paramsdesc><paramgroups>0</paramgroups><rettype>`dict[str, list]`</rettype><retdesc>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.</retdesc></docstring>

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.







<ExampleCodeBlock anchor="trl.maybe_extract_prompt.example">

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.'}]}
```

</ExampleCodeBlock>

<ExampleCodeBlock anchor="trl.maybe_extract_prompt.example-2">

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.'}]}
```

</ExampleCodeBlock>


</div>

## unpair_preference_dataset[[trl.unpair_preference_dataset]]

<div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8">


<docstring><name>trl.unpair_preference_dataset</name><anchor>trl.unpair_preference_dataset</anchor><source>https://github.com/huggingface/trl/blob/vr_4305/trl/data_utils.py#L324</source><parameters>[{"name": "dataset", "val": ": ~DatasetType"}, {"name": "num_proc", "val": ": typing.Optional[int] = None"}, {"name": "desc", "val": ": typing.Optional[str] = None"}]</parameters><paramsdesc>- **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.</paramsdesc><paramgroups>0</paramgroups><rettype>[Dataset](https://huggingface.co/docs/datasets/main/en/package_reference/main_classes#datasets.Dataset)</rettype><retdesc>The unpaired preference dataset.</retdesc></docstring>

Unpair a preference dataset.







<ExampleCodeBlock anchor="trl.unpair_preference_dataset.example">

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}
```

</ExampleCodeBlock>


</div>

## maybe_unpair_preference_dataset[[trl.maybe_unpair_preference_dataset]]

<div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8">


<docstring><name>trl.maybe_unpair_preference_dataset</name><anchor>trl.maybe_unpair_preference_dataset</anchor><source>https://github.com/huggingface/trl/blob/vr_4305/trl/data_utils.py#L367</source><parameters>[{"name": "dataset", "val": ": ~DatasetType"}, {"name": "num_proc", "val": ": typing.Optional[int] = None"}, {"name": "desc", "val": ": typing.Optional[str] = None"}]</parameters><paramsdesc>- **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.</paramsdesc><paramgroups>0</paramgroups><rettype>[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)</rettype><retdesc>The unpaired preference dataset if it was paired, otherwise
the original dataset.</retdesc></docstring>

Unpair a preference dataset if it is paired.







<ExampleCodeBlock anchor="trl.maybe_unpair_preference_dataset.example">

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}
```

</ExampleCodeBlock>


</div>

## pack_dataset[[trl.pack_dataset]]

<div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8">


<docstring><name>trl.pack_dataset</name><anchor>trl.pack_dataset</anchor><source>https://github.com/huggingface/trl/blob/vr_4305/trl/data_utils.py#L661</source><parameters>[{"name": "dataset", "val": ": ~DatasetType"}, {"name": "seq_length", "val": ": int"}, {"name": "strategy", "val": ": str = 'bfd'"}, {"name": "map_kwargs", "val": ": typing.Optional[dict[str, typing.Any]] = None"}]</parameters><paramsdesc>- **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. Sequences are never cut in the
    middle.
  - `"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.</paramsdesc><paramgroups>0</paramgroups><rettype>[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)</rettype><retdesc>The dataset with packed sequences. The number of examples
may decrease as sequences are combined.</retdesc></docstring>

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







<ExampleCodeBlock anchor="trl.pack_dataset.example">

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

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

</ExampleCodeBlock>


</div>

## truncate_dataset[[trl.truncate_dataset]]

<div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8">


<docstring><name>trl.truncate_dataset</name><anchor>trl.truncate_dataset</anchor><source>https://github.com/huggingface/trl/blob/vr_4305/trl/data_utils.py#L717</source><parameters>[{"name": "dataset", "val": ": ~DatasetType"}, {"name": "max_length", "val": ": int"}, {"name": "map_kwargs", "val": ": typing.Optional[dict[str, typing.Any]] = None"}]</parameters><paramsdesc>- **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.</paramsdesc><paramgroups>0</paramgroups><rettype>[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)</rettype><retdesc>The dataset with truncated sequences.</retdesc></docstring>

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







<ExampleCodeBlock anchor="trl.truncate_dataset.example">

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]]}
```

</ExampleCodeBlock>


</div>

<EditOnGithub source="https://github.com/huggingface/trl/blob/main/docs/source/data_utils.md" />