Experiments

What experiment tracking is

Every time the agent runs a training job, Meshia automatically logs it as an experiment. Each experiment captures:

• Config: hyperparameters, model architecture, dataset, seed
• Metrics: loss, accuracy, custom metrics, logged per step and per epoch
• Artifacts: checkpoints, plots, output files
• Environment: GPU type, CUDA version, package versions
• Duration: wall time, GPU time, idle time

You don't need to set this up. The agent instruments your training code automatically.

Viewing experiments

Open the Research view to inspect runs for the current session. You'll see the plan and experiment state that the agent has created.

Open a run to inspect its metrics, config, artifacts, and terminal context.

Comparing runs

Select two or more experiments using the checkboxes, then click Compare. You get:

• Side-by-side metric charts (overlay mode)
• Config diff highlighting what changed between runs
• A summary table of final metrics

This is the fastest way to answer "did that hyperparameter change actually help?"

The reproduce button

Every experiment has a Reproduce button. Click it and Meshia creates a new session with the exact same config: same GPU type, packages, code, and hyperparameters. It then runs the experiment again.

Use this to verify results, test on a different GPU, or pick up where you left off.

Logging custom metrics

If the agent's auto-instrumentation doesn't capture a metric you care about, you can log it explicitly in your training code:

from meshia import log_metric





log_metric("bleu_score", 0.342, step=epoch)


log_metric("perplexity", 14.7, step=epoch)

Custom metrics appear alongside auto-captured ones in the experiments panel.