The Playpen is an extension of the clemcore framework. It contains everything you need to get started training and evaluating models in - or from - interaction. It was originally introduced in our EMNLP 2025 paper of the same title. The code for the experiments in the paper can be found here. What you see here has been further evolved considerably, to support a much wider range of experiments and to be much more user friendly.
If you're here just because you find simulating learning from and in interaction interesting just in itself, welcome!, and move on along to the setup instructions... ... because for now, we have temporarily moved these out of the way, to have a landing page for the EMNLP 2026 LM-Playschool Workshop Shared Task:
You have probably come here via the Offical Shared Task Description. (If not, go there and read up on the rules, and register your team if you haven't already.)
The shared task will require you minimally to use the Playpen (the code you find in this repository here) to evaluate the model you want to contribute. So go to the setup instructions now to set this up.
We will post updates here. The best way to communicate with us about matters concerning the shared task (both code-wise, for evaluation or learning, and regarding procedural matters) is via issues in this repository. (Please use the label "shared-task-2026" for procedural question that likely won't result in changes to the code here.) Remember to also look for closed issues before you file a new one.
Alright. Playschool is in session -- happy learning and interacting!