From Talk to Takeoff: harnessing the PyData Berlin presentation as a catalyst for our larger content efforts.

A well-executed conference talk's beauty lies in its immediate impact and ability to fuel a larger momentum across content, education, and user adoption. When I gave a talk at PyData Berlin with the brilliant Nikolai Liubimovik, I couldn’t imagine its larger impact on our efforts. Let's dive in to see how one talk about Reinforcement Learning with Human Feedback (RLHF) affected our efforts.

Our presentation — an exploration into RLHF — was crafted to inspire and delight our audience with both theoretical understanding and technical code samples (including a live code demonstration). In the talk, we demonstrated how it is possible to fine-tune large language models like GPT-2 using Label Studio within a CoLab notebook.

This would not have been possible without working as a team; Nikolai, Jimmy Whitaker, and I produced and wrote the talk together and created and contributed to a repo of assets and resources for anyone who wanted to get started on their own time. As we unveiled our insights, we were not merely talking about concepts but priming the pump for an influx of user engagement and activation.

The aftermath of the talk was akin to a snowball rolling down a hill, gathering speed and mass with each turn. We witnessed one of the highest spikes in user adoption of Label Studio after the conference. We leaned into this momentum and strategically facilitated an expansion of the talk in the form of a workshop afterward.

We realized that each conference presentation, each insightful conversation, is a jolt of energy that can be amplified into ongoing activations.

The success of PyData Berlin and many other conference presentations and tutorials from our team helped us establish a Label Studio workshop series, providing an arena for continuous interaction and learning. This extended interaction and learning often drove more users to Label Studio.

These sessions were not merely passive one-way communications; they were vibrant exchanges of ideas, doubts, and experiences that propelled our users' journeys with Label Studio. Some of these ideas and interactions led to further product development with Label Studio — such as the new Generative AI Template library.

The workshops, in turn, became an engine that further fueled our content strategy. They provided fresh insights into user needs and feedback, which we integrated into our subsequent content pieces, thus spinning the content wheel.

This strategy, however, was more than a cyclical process of content creation. It weaved together the diverse threads of user engagement, community interaction, and continuous learning into a robust tapestry of knowledge and innovation.

One thing I’ve learned being in AI/ML over the last year is identifying a gap for the middle-of-the-journey folks. There’s content on how to get started and content for the advanced developments in model training and development, but that middle-of-the-journey audience is somewhat neglected. I believe this conference talk, workshop, and related assets were so successful because it filled this accessibility gap for users and potential users.

Our talk at PyData Berlin was not a solitary event but a pivotal piece in our larger content strategy. It was the first turn of the wheel that set the motion for a chain of activations — tutorials, workshops, community interactions, and more.

In the grand scheme of things, every conference talk, every piece of content, and every workshop is a nudge that keeps your flywheel of content strategy turning. As developer advocates or educators, we're excited to keep this wheel in motion, generating and harnessing momentum for continuous learning, growth, and innovation.


Project Recap

Primary Objective: Increase awareness of Label Studio’s ability to help fine-tune models and be seen as a core player in the open source machine learning ecosystem.

Secondary Objective Increase adoption of Label Studio through understanding RLHF.

Results:

  • A significant spike in user adoption after PyData Berlin
  • Evergreen assets that are created and used long after the talk
  • Mentioned as one of the “best talks” at PyData Berlin

Skills used:

  • Python
  • GPT2 API
  • CoLab Notebook/Jupyter Notebook
  • RLHF