Weights & Biases (W&B) — AI Glossary

Reference details

Related terms

• MLflow
• Experiment Tracking
• LangSmith

How it works

1. Each training or evaluation run logs parameters, metrics and artefacts against a run identifier, so an experiment is queryable rather than remembered.
2. Runs group into projects and can be compared on recorded numbers, which is what makes two configurations genuinely comparable.
3. Hyperparameter sweeps are orchestrated by the platform, which proposes configurations and records results per trial.
4. Artefacts version datasets and models with lineage, so a result can be traced back to the exact data that produced it.
5. A separate tracing product extends the same run-centred model to model-backed applications, recording per-request traces and evaluations.

Example: What lineage answers that logging alone does not

A model performs worse in production than in evaluation. Metrics alone show the gap and not its cause. Artefact lineage shows the deployed checkpoint came from a run whose training dataset version predates a cleaning fix applied two weeks earlier — so the evaluation used clean data and the training did not. Parameters and metrics were logged throughout and could not have revealed that; the dataset version link is what made the question answerable.

Commonly confused with

Grouped with LLM observability platforms, which overlap at the edges and come from different directions. This is built around training runs, sweeps and artefact lineage; LLM-native platforms are built around per-request traces, prompt versions and judge scores. A team training or fine-tuning models usually needs the first, a team calling hosted APIs usually needs the second, and a team doing both runs both. It is also not a model registry in the governance sense, though artefacts provide versioning — stage-gated approval is a different product concern.

When to use it

Reach for it when:

• Training or fine-tuning, where runs, sweeps and metric comparison are the workflow.
• Dataset and model lineage must be traceable from a result back to its inputs.
• Hyperparameter search should be orchestrated rather than scripted.

Reach for something else when:

• Only hosted APIs are called, where there are no training runs to track.
• Per-request tracing of an application is the primary need.
• A stage-gated approval lifecycle is required, which is a registry concern.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Weights & Biases (W&B) — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path