Pandas

Pandas is a widely used Python data analysis toolkit. Since it uses fsspec to read and write remote data, you can use the Hugging Face paths (hf://) to read and write data on the Hub.

Load a DataFrame

You can load data from local files or from remote storage like Hugging Face Datasets. Pandas supports many formats including CSV, JSON and Parquet:

>>> import pandas as pd
>>> df = pd.read_csv("path/to/data.csv")

To load a file from Hugging Face, the path needs to start with hf://. For example, the path to the stanfordnlp/imdb dataset repository is hf://datasets/stanfordnlp/imdb. The dataset on Hugging Face contains multiple Parquet files. The Parquet file format is designed to make reading and writing data frames efficient, and to make sharing data across data analysis languages easy. Here is how to load the file plain_text/train-00000-of-00001.parquet:

>>> import pandas as pd
>>> df = pd.read_parquet("hf://datasets/stanfordnlp/imdb/plain_text/train-00000-of-00001.parquet")
>>> df
                                                    text  label
0      I rented I AM CURIOUS-YELLOW from my video sto...      0
1      "I Am Curious: Yellow" is a risible and preten...      0
2      If only to avoid making this type of film in t...      0
3      This film was probably inspired by Godard's Ma...      0
4      Oh, brother...after hearing about this ridicul...      0
...                                                  ...    ...
24995  A hit at the time but now better categorised a...      1
24996  I love this movie like no other. Another time ...      1
24997  This film and it's sequel Barry Mckenzie holds...      1
24998  'The Adventures Of Barry McKenzie' started lif...      1
24999  The story centers around Barry McKenzie who mu...      1

For more information on the Hugging Face paths and how they are implemented, please refer to the the client library’s documentation on the HfFileSystem.

Save a DataFrame

You can save a pandas DataFrame using to_csv/to_json/to_parquet to a local file or to Hugging Face directly.

To save the DataFrame on Hugging Face, you first need to Login with your Hugging Face account, for example using:

hf auth login

Then you can Create a dataset repository, for example using:

from huggingface_hub import HfApi

HfApi().create_repo(repo_id="username/my_dataset", repo_type="dataset")

Finally, you can use Hugging Face paths in Pandas:

import pandas as pd

df.to_parquet("hf://datasets/username/my_dataset/imdb.parquet")

# or write in separate files if the dataset has train/validation/test splits
df_train.to_parquet("hf://datasets/username/my_dataset/train.parquet")
df_valid.to_parquet("hf://datasets/username/my_dataset/validation.parquet")
df_test .to_parquet("hf://datasets/username/my_dataset/test.parquet")

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

• Parquet Content Defined Chunking optimizes Parquet for Xet, 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

Pandas require extra argument to write optimized Parquet files:

import pandas as pd

df.to_parquet(
    "hf://datasets/username/my_dataset/imdb.parquet",
    # Optimize for Xet
    use_content_defined_chunking=True,
    write_page_index=True,
)

• use_content_defined_chunking=True to enable Parquet Content Defined Chunking, for deduplication and editing (it requires pyarrow>=21.0)
• write_page_index=True to include a page index in the Parquet metadata, for streaming and random access

Content defined chunking (CDC) makes the Parquet writer chunk the data pages in a way that makes duplicate data chunked and compressed identically. Without CDC, the pages are arbitrarily chunked and therefore duplicate data are impossible to detect because of compression. Thanks to CDC, Parquet uploads and downloads from Hugging Face are faster, since duplicate data are uploaded or downloaded only once.

Find more information about Xet here.

Leverage Xet deduplication for Parquet

Optimized Parquet files are written with Content Defined Chunking, which enables deduplication. This accelerates uploads since chunks of data that already exist on Hugging Face don’t need to be uploaded again, and this saves a lot of I/O.

For example, this code uploads the content of df and then for edited_df the upload is faster since it only uploads the chunks that changed:

import pandas as pd

df.to_parquet(
    "hf://datasets/username/my_dataset/imdb.parquet",
    # Optimize for Xet
    use_content_defined_chunking=True,
    write_page_index=True,
)

edited_df = ...  # e.g. with added/modified/removed rows or columns

edited_df.to_parquet(
    "hf://datasets/username/my_dataset/imdb.parquet",
    # Optimize for Xet
    use_content_defined_chunking=True,
    write_page_index=True,
)

Chunks are ~64kB and Parquet saves data column per column, so in practice this is what happens when editing an Optimized Parquet file:

• add a new column -> only the chunks of the new column are uploaded
• add/edit/delete a row -> one chunk per column is uploaded

And in addition to this, the chunks of the Parquet footer containing metadata are also uploaded.

Use Images

You can load a folder with a metadata file containing a field for the names or paths to the images, structured like this:

Example 1:            Example 2:
folder/               folder/
├── metadata.csv      ├── metadata.csv
├── img000.png        └── images
├── img001.png            ├── img000.png
...                       ...
└── imgNNN.png            └── imgNNN.png

You can iterate on the images paths like this:

import pandas as pd

folder_path = "path/to/folder/"
df = pd.read_csv(folder_path + "metadata.csv")
for image_path in (folder_path + df["file_name"]):
    ...

Since the dataset is in a supported structure (a metadata.csv or .jsonl file with a file_name field), you can save this dataset to Hugging Face and the Dataset Viewer shows both the metadata and images on Hugging Face.

from huggingface_hub import HfApi
api = HfApi()

api.upload_folder(
    folder_path=folder_path,
    repo_id="username/my_image_dataset",
    repo_type="dataset",
)

Image methods and Parquet

Using pandas-image-methods you enable PIL.Image methods on an image column. It also enables saving the dataset as one single Parquet file containing both the images and the metadata:

import pandas as pd
from pandas_image_methods import PILMethods

pd.api.extensions.register_series_accessor("pil")(PILMethods)

df["image"] = (folder_path + df["file_name"]).pil.open()
df.to_parquet("data.parquet")

All the PIL.Image methods are available, e.g.

df["image"] = df["image"].pil.rotate(90)

Use Audios

You can load a folder with a metadata file containing a field for the names or paths to the audios, structured like this:

Example 1:            Example 2:
folder/               folder/
├── metadata.csv      ├── metadata.csv
├── rec000.wav        └── audios
├── rec001.wav            ├── rec000.wav
...                       ...
└── recNNN.wav            └── recNNN.wav

You can iterate on the audios paths like this:

import pandas as pd

folder_path = "path/to/folder/"
df = pd.read_csv(folder_path + "metadata.csv")
for audio_path in (folder_path + df["file_name"]):
    ...

Since the dataset is in a supported structure (a metadata.csv or .jsonl file with a file_name field), you can save it to Hugging Face, and the Hub Dataset Viewer shows both the metadata and audio.

from huggingface_hub import HfApi
api = HfApi()

api.upload_folder(
    folder_path=folder_path,
    repo_id="username/my_audio_dataset",
    repo_type="dataset",
)

Audio methods and Parquet

Using pandas-audio-methods you enable soundfile methods on an audio column. It also enables saving the dataset as one single Parquet file containing both the audios and the metadata:

import pandas as pd
from pandas_audio_methods import SFMethods

pd.api.extensions.register_series_accessor("sf")(SFMethods)

df["audio"] = (folder_path + df["file_name"]).sf.open()
df.to_parquet("data.parquet")

This makes it easy to use with librosa e.g. for resampling:

df["audio"] = [librosa.load(audio, sr=16_000) for audio in df["audio"]]
df["audio"] = df["audio"].sf.write()

Use Transformers

You can use transformers pipelines on pandas DataFrames to classify, generate text, images, etc. This section shows a few examples with tqdm for progress bars.

Pipelines don’t accept a tqdm object as input but you can use a python generator instead, in the form x for x in tqdm(...)

Text Classification

from transformers import pipeline
from tqdm import tqdm

pipe = pipeline("text-classification", model="clapAI/modernBERT-base-multilingual-sentiment")

# Compute labels
df["label"] = [y["label"] for y in pipe(x for x in tqdm(df["text"]))]
# Compute labels and scores
df[["label", "score"]] = [(y["label"], y["score"]) for y in pipe(x for x in tqdm(df["text"]))]

Text Generation

from transformers import pipeline
from tqdm import tqdm

pipe = pipeline("text-generation", model="Qwen/Qwen2.5-1.5B-Instruct")

# Generate chat response
prompt = "What is the main topic of this sentence ? REPLY IN LESS THAN 3 WORDS. Sentence: '{}'"
df["output"] = [y["generated_text"][1]["content"] for y in pipe([{"role": "user", "content": prompt.format(x)}] for x in tqdm(df["text"]))]