What is WandB?
Weights & Biases or WandB is a machine learning platform that covers the full model development lifecycle — from experiment tracking and training to deployment monitoring and LLM observability. It gives ML teams a single environment to build, evaluate, and ship AI models and agentic applications.
The platform splits into two core product surfaces. W&B Models handles experiment tracking, hyperparameter optimization, dataset versioning, and model registry. W&B Weave handles LLM application development, including trace logging, prompt evaluation, guardrails, and production monitoring. Both surfaces share a common registry, SDK, and automation layer.
Features & Benefits
- Experiment Tracking — Log, visualize, and compare training runs across metrics, hyperparameters, and system performance in real time.
- Hyperparameter Optimization — Run automated hyperparameter sweeps to find optimal model configurations without manual tuning.
- Dataset and Artifact Versioning — Version datasets, models, prompts, and code with full lineage tracking across every experiment.
- Model Registry — Publish, organize, and share trained models across teams with metadata, tags, and lifecycle stage tracking.
- Serverless Fine-Tuning — Run supervised fine-tuning and reinforcement learning jobs on hosted GPU infrastructure without managing compute.
- Reinforcement Learning Framework — Build and run RL training pipelines using ART, an open-source framework, with Ruler for automated reward function generation.
- Hosted Model Inference — Access and serve hosted open-source models including Llama, DeepSeek, Qwen, and others directly through the platform.
- LLM Trace Logging — Capture every LLM call, retrieval step, and agent action for debugging and performance analysis.
- LLM Evaluation — Run structured, repeatable evaluations of LLM outputs and agentic pipelines against defined criteria.
- Prompt Playground — Test and compare prompts across models interactively before deploying to production.
- Agent Observability — Monitor agentic systems with trace-level visibility into multi-step reasoning and tool use.
- Guardrails — Block prompt injection attacks and filter harmful or out-of-scope model outputs before they reach end users.
- Production Monitors — Set up continuous evaluations to detect performance regressions in deployed models and LLM applications.
- Collaborative Reports — Build shareable dashboards that document experiments, findings, and model comparisons.
- Workflow Automations — Trigger downstream actions automatically based on registry events or training milestones.
- Integrations — Connect with major ML frameworks and cloud providers through native SDK support.
Real-World Applications
Teams training large models on PyTorch, TensorFlow, or Hugging Face Transformers can connect WandB with a few lines of code. Every run logs metrics, gradients, and system stats automatically. Researchers can compare dozens of runs side by side and trace any result back to the exact data and configuration that produced it.
Organizations building RAG pipelines or LLM-powered applications may use Weave to capture full execution traces across retrieval, generation, and response steps. Evaluation tools let teams score outputs systematically and track quality changes as prompts or models are updated. Guardrails can block unsafe or off-topic responses before they surface to end users.
Computer vision and physical AI teams working on tasks like autonomous driving or robotics can use the dataset versioning and artifact tracking tools to manage large-scale training data. Model checkpoints, annotations, and evaluation results stay linked to the runs that produced them, making it easier to audit and reproduce results across long research cycles.
Enterprise teams that need to fine-tune LLMs for domain-specific tasks — such as contact center automation or financial document analysis — can run serverless supervised fine-tuning without provisioning or managing GPU infrastructure. The machine learning platform handles compute allocation, and trained models feed directly into the Registry for versioning and deployment.
