Weights & Biases

OpenResearch automatically detects W&B runs started during your experiment and links them in the UI. No configuration or code changes required.

How it works

When your run script calls wandb.init(), OpenResearch captures the run URL automatically. A W&B link appears next to the run's status badge, updating in real time as new W&B runs are created.

This works with any setup: plain Python, uv, virtualenvs, shell scripts that call Python. Multiple wandb.init() calls in a single run (e.g. separate training and evaluation runs) are all captured.

Authentication

W&B needs a WANDB_API_KEY to log runs. You can find yours at wandb.ai/authorize. Add it as an environment variable in your org or project settings. Org-level env vars apply to all projects, so if your whole team uses the same W&B account, set it once at the org level.

You can also set WANDB_PROJECT and WANDB_ENTITY to control where runs are logged without hardcoding them in your scripts.

Self-hosted W&B

If you're using a self-hosted W&B instance, the links will automatically point to your server. OpenResearch reads the URL directly from the W&B run object, so no extra configuration is needed.