# Libraries

The Datasets Hub has support for several libraries in the Open Source ecosystem.
Thanks to the [huggingface_hub Python library](/docs/huggingface_hub), it's easy to enable sharing your datasets on the Hub.
We're happy to welcome to the Hub a set of Open Source libraries that are pushing Machine Learning forward.

## Libraries table

The table below summarizes the supported libraries and their level of integration.

| Library                             | Description                                                                                                                    | Download from Hub | Stream from Hub | Push to Hub | Stream to Hub | Optimized Parquet files |
| ----------------------------------- | ------------------------------------------------------------------------------------------------------------------------------ | ----------------- | --------------- | ----------- | ------------- | ----------------------- |
| [Argilla](./datasets-argilla)       | Collaboration tool for AI engineers and domain experts that value high quality data.                                           | ✅                | ❌              | ✅          | ❌            | ❌                      |
| [Daft](./datasets-daft)             | Data engine for large scale, multimodal data processing with a Python-native interface.                                        | ✅                | ✅              | ✅          | ✅            | ✅                      |
| [Dask](./datasets-dask)             | Parallel and distributed computing library that scales the existing Python and PyData ecosystem.                               | ✅                | ✅              | ✅          | ✅            | ✅*                     |
| [Data Designer](./datasets-data-designer) | NVIDIA NeMo framework for generating synthetic datasets using LLMs.                                                      | ✅                | ❌              | ✅          | ❌            | ❌                      |
| [Datasets](./datasets-usage)        | 🤗 Datasets is a library for accessing and sharing datasets for Audio, Computer Vision, and Natural Language Processing (NLP). | ✅                | ✅              | ✅          | ✅            | ✅                      |
| [Distilabel](./datasets-distilabel) | The framework for synthetic data generation and AI feedback.                                                                   | ✅                | ❌              | ✅          | ❌            | ❌                      |
| [DuckDB](./datasets-duckdb)         | In-process SQL OLAP database management system.                                                                                | ✅                | ✅              | ❌          | ❌            | ❌                      |
| [Embedding Atlas](./datasets-embedding-atlas) | Interactive visualization and exploration tool for large embeddings.                                                 | ✅                | ✅              | ❌          | ❌            | ❌                      |
| [Fenic](./datasets-fenic)           | PySpark-inspired DataFrame framework for building production AI and agentic applications.                                      | ✅                | ✅              | ❌          | ❌            | ❌                      |
| [FiftyOne](./datasets-fiftyone)     | FiftyOne is a library for curation and visualization of image, video, and 3D data.                                             | ✅                | ✅              | ✅          | ❌            | ❌                      |
| [Lance](./datasets-lance)           | An open lakehouse format for multimodal AI.                                                                                    | ✅                | ✅              | ❌          | ❌            | ❌                      |
| [Pandas](./datasets-pandas)         | Python data analysis toolkit.                                                                                                  | ✅                | ❌              | ✅          | ❌            | ✅*                     |
| [Polars](./datasets-polars)         | A DataFrame library on top of an OLAP query engine.                                                                            | ✅                | ✅              | ✅          | ❌            | ❌                      |
| [PyArrow](./datasets-pyarrow)       | Apache Arrow is a columnar format and a toolbox for fast data interchange and in-memory analytics.                             | ✅                | ✅              | ✅          | ❌            | ✅*                     |
| [Spark](./datasets-spark)           | Real-time, large-scale data processing tool in a distributed environment.                                                      | ✅                | ✅              | ✅          | ✅            | ✅                      |
| [WebDataset](./datasets-webdataset) | Library to write I/O pipelines for large datasets.                                                                             | ✅                | ✅              | ❌          | ❌            | ❌                      |

_ * Requires passing extra arguments to write optimized Parquet files_

## Data Processing Libraries

### Streaming

Dataset streaming allows iterating on a dataset from Hugging Face progressively without having to download it completely.
It saves local disk space because the data is never on disk. It saves memory since only a small portion of the dataset is used at a time. And it saves time, since there is no need to download data before the CPU or GPU workload.

In addition to streaming *from* Hugging Face, many libraries also support streaming *back to* Hugging Face.
Therefore, they can run end-to-end streaming pipelines: streaming from a source and writing to Hugging Face progressively, often overlapping the download, upload, and processing steps.

For more details on how to do streaming, check out the documentation of a library that support streaming (see table above) or the [streaming datasets](./datasets-streaming) documentation if you want to stream datasets from Hugging Face by yourself.

### Optimized Parquet files

Parquet files on Hugging Face are optimized to improve storage efficiency, accelerate downloads and uploads, and enable efficient dataset streaming and editing.

Optimized Parquet files are Parquet files with additional features:

* [Parquet Content Defined Chunking](https://huggingface.co/blog/parquet-cdc) optimizes Parquet for [Xet](https://huggingface.co/docs/hub/en/xet/index), Hugging Face's storage backend. It accelerates uploads and downloads thanks to chunk-based deduplication and allows efficient file editing
* Page index accelerates filters when streaming and enables efficient random access, e.g. in the [Dataset Viewer](https://huggingface.co/docs/dataset-viewer)

Some libraries require extra argument to write Optimized Parquet files like `Pandas` and `PyArrow`:

* `use_content_defined_chunking=True` to enable Parquet Content Defined Chunking, for [deduplication](https://huggingface.co/blog/parquet-cdc) and [editing](./datasets-editing)
* `write_page_index=True` to include a page index in the Parquet metadata, for [streaming and random access](./datasets-streaming)

## Training Libraries

Training libraries that integrate with Hub datasets for model training. The table below shows their streaming capabilities - the ability to train on datasets without downloading them first.

| Library | Description | Stream from Hub |
| ------- | ----------- | --------------- |
| [Axolotl](https://docs.axolotl.ai/docs/streaming.html) | Low-code LLM fine-tuning framework | ✅ |
| [LlamaFactory](https://github.com/hiyouga/LLaMA-Factory) | Unified fine-tuning for 100+ LLMs | ✅ |
| [Sentence Transformers](https://sbert.net/docs/sentence_transformer/training_overview.html) | Text embeddings and semantic similarity | ✅ |
| [Transformers](https://huggingface.co/docs/transformers/trainer) | 🤗 Transformers Trainer for fine-tuning models | ✅ |
| [TRL](https://huggingface.co/docs/trl) | Training LLMs with reinforcement learning (SFT, DPO, GRPO) | ⚠️* |
| [Unsloth](https://docs.unsloth.ai) | Fast LLM fine-tuning (2x speedup, 70% less memory) | ✅ |

_* SFTTrainer and DPOTrainer support streaming; GRPOTrainer does not yet support streaming input_

### Streaming from Hub

Streaming allows training on massive datasets without downloading them first. This is valuable when:
- Your dataset is too large to fit on disk
- You want to start training immediately
- You're using [HF Jobs](https://huggingface.co/docs/hub/jobs) where co-located compute provides faster streaming

Recent improvements have made streaming [up to 100x more efficient](https://huggingface.co/blog/streaming-datasets) with faster startup, prefetching, and better scaling to many workers.

**Note:** Streaming requires `max_steps` in training arguments since dataset length is unknown, and uses buffer-based shuffling. See [streaming datasets](./datasets-streaming) for more details.

### Logging to Hub

Some tools can stream training data back to the Hub during training:

- **[Trackio](https://github.com/huggingface/trackio)**: Streams training metrics to a Hub dataset in real-time

## Integrating data libraries and tools with the Hub

This guide is designed for developers and maintainers of data libraries and tools who want to integrate with the Hugging Face Hub. Whether you're building a data processing library, analysis tool, or any software that needs to interact with datasets, this documentation will help you implement a Hub integration.

The guide covers:

- Possible approaches to loading data from the Hub into your library/tool
- Possible approaches to uploading data from your library/tool to the Hub

### Loading data from the Hub

If you have a library for working with data, it can be helpful for your users to load data from the Hub.

In general, we suggest relying on an existing library like `datasets`, `pandas` or `polars` to do this unless you have a specific reason to implement your own. If you require more control over the loading process, you can use the `huggingface_hub` library, which will allow you, for example, to download a specific subset of files from a repository.

You can find more information about loading data from the Hub [here](https://huggingface.co/docs/hub/datasets-downloading).

#### Integrating via the Dataset Viewer and Parquet Files

The Hub's dataset viewer and Parquet conversion system provide a standardized way to integrate with datasets, regardless of their original format. This infrastructure is a reliable integration layer between the Hub and external libraries.

If the dataset is not already in Parquet, the Hub automatically converts the first 5GB of every dataset to Parquet format to power the dataset viewer and provide consistent access patterns. This standardization offers several benefits for library integrations:

- Consistent data access patterns regardless of original format
- Built-in dataset preview and exploration through the Hub's dataset viewer. The dataset viewer can also be embedded as an iframe in your applications, making it easy to provide rich dataset previews. For more information about embedding the viewer, see the [dataset viewer embedding documentation](https://huggingface.co/docs/hub/en/datasets-viewer-embed).
- Efficient columnar storage optimized for querying. For example, you could use a tool like [DuckDB](https://duckdb.org/) to query or filter for a specific subset of data.
- Parquet is well supported across the machine learning and data science ecosystem.

For more details on working with the Dataset Viewer API, see the [Dataset Viewer API documentation](https://huggingface.co/docs/dataset-viewer/index)

### Uploading data to the Hub

This section covers possible approaches for adding the ability to upload data to the Hub in your library, i.e. how to implement a `push_to_hub` method.

This guide will cover three primary ways to upload data to the Hub:

- using the `datasets` library and the `push_to_hub` method
- using `pandas` to write to the Hub
- using the `huggingface_hub` library and the `hf_hub_download` method
- directly using the API or Git with git-xet

#### Use the `datasets` library

The most straightforward approach to pushing data to the Hub is to rely on the existing [`push_to_hub`](https://huggingface.co/docs/datasets/main/en/package_reference/main_classes#datasets.Dataset.push_to_hub) method from the `datasets` library. The `push_to_hub` method will automatically handle:

- the creation of the repository
- the conversion of the dataset to Parquet
- chunking the dataset into suitable parts
- uploading the data

For example, if you have a synthetic data generation library that returns a list of dictionaries, you could simply do the following:

```python
from datasets import Dataset

data = [{"prompt": "Write a cake recipe", "response": "Measure 1 cup ..."}]
ds = Dataset.from_list(data)
ds.push_to_hub("USERNAME_OR_ORG/repo_ID")
```

Examples of this kind of integration:

- [Distilabel](https://github.com/argilla-io/distilabel/blob/8ad48387dfa4d7bd5639065661f1975dcb44c16a/src/distilabel/distiset.py#L77)

#### Rely on an existing libraries integration with the Hub

Polars, Pandas, Dask, Spark, DuckDB, and Daft can all write to a Hugging Face Hub repository. See [datasets libraries](https://huggingface.co/docs/hub/datasets-libraries) for more details.

If you are already using one of these libraries in your code, adding the ability to push to the Hub is straightforward. For example, if you have a synthetic data generation library that can return a Pandas DataFrame, here is the code you would need to write to the Hub:

```python
from huggingface_hub import HfApi

# Initialize the Hub API
hf_api = HfApi(token=os.getenv("HF_TOKEN"))

# Create a repository (if it doesn't exist)
hf_api.create_repo(repo_id="username/my-dataset", repo_type="dataset")

# Convert your data to a DataFrame and save directly to the Hub
df.to_parquet("hf://datasets/username/my-dataset/data.parquet")
```

#### Using the huggingface_hub Python library

The `huggingface_hub` Python library offers a more flexible approach to uploading data to the Hub. The library allows you to upload specific files or subsets of files to a repository. This is useful if you have a large dataset that you don't want to convert to Parquet, want to upload a specific subset of files, or want more control over the repo structure.

Depending on your use case, you can upload a file or folder at a specific point in your code, i.e., export annotations from a tool to the Hub when a user clicks "push to Hub". For example,

```python
from huggingface_hub import HfApi
api = HfApi(token=HF_TOKEN)

api.upload_folder(
    folder_path="/my-cool-library/data-folder",
    repo_id="username/my-cool-space",
    repo_type="dataset",
    commit_message="Push annotations to Hub"
    allow_patterns="*.jsonl",
)
```

You can find more information about ways to upload data to the Hub [here](https://huggingface.co/docs/huggingface_hub/main/en/guides/upload).

Alternatively, there are situations where you may want to upload data in the background, for example, synthetic data being generated every 10 minutes. In this case you can use the `scheduled_uploads` feature of the `huggingface_hub` library. For more details, see the [scheduled uploads documentation](https://huggingface.co/docs/huggingface_hub/main/en/guides/upload#scheduled-uploads).

You can see an example of using this approach to upload data to the Hub in

- The [fastdata](https://github.com/AnswerDotAI/fastdata/blob/main/nbs/00_core.ipynb) library
- This [magpie](https://huggingface.co/spaces/davanstrien/magpie/blob/fc79672c740b8d3d098378dca37c0f191c208de0/app.py#L67) Demo Space

## More support

For technical questions about integration, feel free to contact the datasets team at datasets@huggingface.co.