Weight and bias

I am training an RL algorithm for LLM post-training and would like to track the training process using Weights & Biases.

However, in my company environment, I am not allowed to send any data to external servers for security reasons. Because of this, I tried using W&B in offline mode.

Here is how I initialize it:

wandb.init(mode="offline", project="PPO", name="LLM_run_1")


Unfortunately, this setup does not seem to work as expected, and I am unable to properly track or visualize my runs.

Could you please clarify:

  • How to correctly use W&B in a fully offline environment?

  • Whether there is a way to visualize runs locally without syncing to the cloud?

  • Or if there are recommended best practices for this use case?

Any guidance would be appreciated. Thank you!