Thomas Wolf

Who he is

Thomas Wolf is a co-founder and chief science officer of Hugging Face, the company he started in 2016 together with Clément Delangue and Julien Chaumond^[1]. His title places him on the scientific side of the organisation: the co-founders handle the company, while the chief science officer is associated with the research and engineering work that the company publishes and maintains.

The Transformers paper

Wolf's most visible piece of published work is the paper "HuggingFace's Transformers: State-of-the-art Natural Language Processing", which describes the company's Transformers library^[2]. He is the first author on it, listed ahead of Lysandre Debut, Victor Sanh and Julien Chaumond among a total of 22 authors^[3]. The paper was submitted to arXiv on 9 October 2019 and was revised for the last time on 14 July 2020^[3], so it was written while the library was already in use and continued to be updated as the library changed.

What the library is

In the paper, Transformers is described as an open-source library that offers state-of-the-art Transformer architectures under a unified API, backed by a curated collection of pretrained models^[2]. The phrase "unified API" is the key one. A Transformer is a neural network architecture for handling sequences of text; many research groups produced their own versions of it, each with its own code conventions, file formats and loading routines. The library's contribution was to put those architectures behind one common interface, so that switching from one model to another did not mean rewriting a project.

The "curated collection of pretrained models" matters for the same reason. Training such a model from scratch requires a large amount of text and computing time. A pretrained model can instead be downloaded and adapted to a new task with far less work, and a curated collection means the models are checked and packaged rather than scattered across the web.

Design goals

According to the same paper, the library was designed to be extensible by researchers, simple for practitioners and fast and robust in industrial deployments^[2]. Those three goals pull in different directions: researchers want to add new architectures and modify internals, practitioners want to load a model and run it without reading the source code, and industrial users care about speed and reliability once the model is in production. Stating all three as targets shows that the library was not aimed at a single audience.

Why it matters

The paper is a description of a tool rather than a report of a new model or a benchmark result. Its significance lies in the tool: when a shared library becomes the common way to obtain and run pretrained language models, results become easier to reproduce and compare, and people who are not machine-learning specialists can use the models at all. Wolf's position as first author of that paper, and as co-founder and chief science officer of the company that maintains the library, is the main public record of his work.