Module 06 · Integrations

Weights & Biases

Let the Coder and the Scientist read your training runs, metrics, comparisons, artifacts, and traces in Weights & Biases. Read only.

View as Markdown

On this page

Weights & Biases is where many teams track machine learning experiments: loss curves, metrics, configs, saved artifacts, and model registries. Once you connect it, the Coder and the Scientist can read your runs, compare them, and tell you which result to trust. They never write to your account.

Which assistants use it#

What assistants can do#

Weights & Biases connects through its own hosted connector, and it is read only.

Ability	How it runs
List your teams and projects	Reads on its own
Read a run's metrics and loss curve	Reads on its own
Compare runs and diagnose a run	Reads on its own
Read artifacts, versions, and model registries	Reads on its own
Read evaluation results and traces	Reads on its own
Search the Weights & Biases documentation	Reads on its own
Create a report or log a run	Never run from chat
Send a raw query that could change something, or start the vendor's own agent	Never run from chat

Setup#

In Weights & Biases, open wandb.ai/authorize and copy your key. For a shared team, create a service account key limited to the team whose runs the assistant should read.
Open Integrations in StellarFirm and pick Weights & Biases.
Paste the key, switch on the live option, and press Connect.

Permissions#

Nothing here waits for an Approve because nothing here changes data.
Revoke the key in Weights & Biases whenever you want to cut access.

Good to know#

Training runs on a StellarFirm assistant's own computer do not reach your Weights & Biases account: those computers have no GPU and run only a small proof run, offline. The real run happens on a GPU you control, which you start from the Coder's RUN.md, and its results appear here for the assistant to read.
Set your key only on the machine that runs the training, never in the chat or in a file in your repository.
Every figure comes with its team, project, and run id, so you can open the same run yourself.
Run names, notes, and trace text are outside information. An assistant reads them as data, never as instructions.

Prompts to try#

PromptWhich run to trust
Coder, list my projects, compare the last three runs of the newest one, and tell me which run to trust and why, with each run's id.

PromptIs the loss healthy
Scientist, read the loss curve and gradient norm of run [run id] and say whether it looks healthy, and what I should check first if it does not.