Beast-logger is a logging kit built for LLM systems, providing reliable high-resolution token-level LLM activity that is unprecedented in any other projects.

Here is how to use beast-logger in AgentJet.

Usage in AgentJet

  1. Start training or debugging with AgentJet launcher.

  2. Wait until the first batch is completed.

  3. Locate log files. By default, they will be placed at saved_experiments/${experiment_name}. For example: saved_experiments/benchmark_frozenlake_20251223_2305

  4. Run beast_logger_go command in the VSCode terminal (or any other software with port-forwarding ability) to start the web log-viewer. Click http://127.0.0.1:8181 to open it (VSCode will automatically forward this port from server -> your local computer)

image
  1. Fill the ABSOLUTE path of the log files and click submit.

    Hint: absolute path is recommended.

    However, you can also use relative path, if beast_logger_go command is launched at same working dir.

    Warning: Beast-logger recursively scans this path,

    thus, where possible, selects the innermost directory containing the fewest files to read logs faster.

image
  1. Choose entry to display

  2. Yellow tokens: tokens that are excluded from loss computation.

  3. Blue tokens: tokens that participant loss computation.
  4. Hovor your mouse on one of the tokens: show the logprob value of that token.
image