Hugging Face Integrations

Hugging Face rapidly became a very popular platform to build, share and collaborate on deep learning applications. We have worked on integrating the torch for R ecosystem with Hugging Face tools, allowing users to load and execute language models from their platform.

We are happy to announce the first releases of hfhub and tok are now on CRAN. hfhub is an R interface to Hugging Face Hub, allowing users to download and cache files from Hugging Face Hub while tok implements R bindings for the Hugging Face tokenizers library.

Hugging Face rapidly became the platform to build, share and collaborate on deep learning applications and we hope these integrations will help R users to get started using Hugging Face tools as well as building novel applications.

We also have previously announced the safetensors package allowing to read and write files in the safetensors format.

hfhub

hfhub is an R interface to the Hugging Face Hub. hfhub currently implements a single functionality: downloading files from Hub repositories. Model Hub repositories are mainly used to store pre-trained model weights together with any other metadata necessary to load the model, such as the hyperparameters configurations and the tokenizer vocabulary.

Downloaded files are ached using the same layout as the Python library, thus cached files can be shared between the R and Python implementation, for easier and quicker switching between languages.

We already use hfhub in the minhub package and in the ‘GPT-2 from scratch with torch’ blog post to download pre-trained weights from Hugging Face Hub.

You can use hub_download() to download any file from a Hugging Face Hub repository by specifying the repository id and the path to file that you want to download. If the file is already in the cache, then the function returns the file path imediately, otherwise the file is downloaded, cached and then the access path is returned.

path <- hfhub::hub_download("gpt2", "model.safetensors")
path
#> /Users/dfalbel/.cache/huggingface/hub/models--gpt2/snapshots/11c5a3d5811f50298f278a704980280950aedb10/model.safetensors

tok

Tokenizers are responsible for converting raw text into the sequence of integers that is often used as the input for NLP models, making them an critical component of the NLP pipelines. If you want a higher level overview of NLP pipelines, you might want to read our previous blog post ‘What are Large Language Models? What are they not?’.

When using a pre-trained model (both for inference or for fine tuning) it’s very important that you use the exact same tokenization process that has been used during training, and the Hugging Face team has done an amazing job making sure that its algorithms match the tokenization strategies used most LLM’s.

tok provides R bindings to the 🤗 tokenizers library. The tokenizers library is itself implemented in Rust for performance and our bindings use the extendr project to help interfacing with R. Using tok we can tokenize text the exact same way most NLP models do, making it easier to load pre-trained models in R as well as sharing our models with the broader NLP community.

tok can be installed from CRAN, and currently it’s usage is restricted to loading tokenizers vocabularies from files. For example, you can load the tokenizer for the GPT2 model with:

tokenizer <- tok::tokenizer$from_pretrained("gpt2")
ids <- tokenizer$encode("Hello world! You can use tokenizers from R")$ids
ids
#> [1] 15496   995     0   921   460   779 11241 11341   422   371
tokenizer$decode(ids)
#> [1] "Hello world! You can use tokenizers from R"

Spaces

Remember that you can already host Shiny (for R and Python) on Hugging Face Spaces. As an example, we have built a Shiny app that uses:

• torch to implement GPT-NeoX (the neural network architecture of StableLM - the model used for chatting)
• hfhub to download and cache pre-trained weights from the StableLM repository
• tok to tokenize and pre-process text as input for the torch model. tok also uses hfhub to download the tokenizer’s vocabulary.

The app is hosted at in this Space. It currently runs on CPU, but you can easily switch the the Docker image if you want to run it on a GPU for faster inference.

The app source code is also open-source and can be found in the Spaces file tab.

Looking forward

It’s the very early days of hfhub and tok and there’s still a lot of work to do and functionality to implement. We hope to get community help to prioritize work, thus, if there’s a feature that you are missing, please open an issue in the GitHub repositories.