weights biases
**Weights & Biases (W&B)** is the **developer-first MLOps platform for experiment tracking, hyperparameter optimization, and model management** — providing real-time, interactive visualizations of training runs that sync to the cloud instantly, enabling ML researchers and engineers to collaborate, compare experiments, and identify what makes models perform better with a UI designed for the way researchers actually work.
**What Is Weights & Biases?**
- **Definition**: A commercial MLOps platform founded in 2017 that provides experiment tracking (Runs), hyperparameter search (Sweeps), dataset and model versioning (Artifacts), and model evaluation tooling — accessed via a Python SDK that integrates with any ML framework and syncs data to W&B's cloud servers in real time.
- **Design Philosophy**: "It just works for researchers" — W&B was designed from the perspective of ML researchers who want to focus on experiments, not infrastructure. Three lines of code to add W&B to any training script; rich visualizations available immediately without configuration.
- **Why W&B Won**: While MLflow focused on enterprise MLOps management, W&B focused on the researcher experience — live loss curves, system metrics (GPU utilization, memory), and one-click sharing of experiment results. This "show your work" culture made W&B viral in research.
- **Enterprise Adoption**: OpenAI, NVIDIA, Samsung, Toyota Research, and hundreds of enterprise ML teams use W&B — the combination of researcher-friendly UX and enterprise features (private cloud, SSO, audit logs) made it the dominant commercial experiment tracking platform.
- **W&B vs MLflow**: W&B is SaaS-first with a polished UI and collaboration features; MLflow is open-source with self-hosting flexibility. W&B excels at research collaboration; MLflow excels at integration with existing enterprise infrastructure.
**Why W&B Matters for AI**
- **Live Training Visualization**: Loss curves, accuracy, learning rate, and custom metrics update in real time as training runs — researchers watch experiments evolve without SSH-ing into training servers to tail log files.
- **System Monitoring**: W&B automatically captures GPU utilization, GPU memory, CPU, RAM, and network metrics — instantly understand if training is GPU-bound, memory-bound, or I/O-bound.
- **Experiment Sharing**: Share a W&B run URL with a colleague or manager — they see the complete experiment: all parameters, metrics charts, system metrics, code, and artifacts in a browser without any setup.
- **Sweeps (HPO)**: W&B Sweeps implements Bayesian optimization, random search, and grid search for hyperparameter tuning — define the search space in a YAML file and W&B launches and manages parallel training runs automatically.
- **Artifacts**: Version control for datasets and models — each dataset version has a hash, lineage to training runs that used it, and downstream model versions that depended on it.
**W&B Core Components**
**Experiment Tracking (Runs)**:
import wandb
wandb.init(
project="llm-fine-tuning",
name="llama-3-8b-lora-v3",
config={
"model": "meta-llama/Llama-3-8B",
"learning_rate": 2e-4,
"lora_rank": 16,
"batch_size": 8,
"epochs": 3
}
)
for epoch in range(config.epochs):
train_loss = train_epoch()
val_loss = evaluate()
wandb.log({
"train/loss": train_loss,
"val/loss": val_loss,
"train/epoch": epoch
})
# Log final model as artifact
wandb.log_artifact("model_checkpoint/", name="fine-tuned-llama", type="model")
wandb.finish()
**Auto-Logging (HuggingFace Integration)**:
from transformers import TrainingArguments
training_args = TrainingArguments(
output_dir="./output",
report_to="wandb", # One flag enables W&B logging
run_name="llama-experiment-v5"
)
# HuggingFace Trainer automatically logs all metrics to W&B
**Sweeps (Hyperparameter Search)**:
sweep_config = {
"method": "bayes", # Bayesian optimization
"metric": {"name": "val/loss", "goal": "minimize"},
"parameters": {
"learning_rate": {"min": 1e-5, "max": 1e-3, "distribution": "log_uniform"},
"lora_rank": {"values": [8, 16, 32, 64]},
"batch_size": {"values": [4, 8, 16]}
}
}
sweep_id = wandb.sweep(sweep_config, project="llm-fine-tuning")
def train():
with wandb.init() as run:
config = run.config
model = train_with_config(config.learning_rate, config.lora_rank)
val_loss = evaluate(model)
wandb.log({"val/loss": val_loss})
wandb.agent(sweep_id, function=train, count=50) # Run 50 experiments
**Artifacts (Data & Model Versioning)**:
# Log dataset as versioned artifact
artifact = wandb.Artifact("training-dataset", type="dataset")
artifact.add_dir("./data/")
run.log_artifact(artifact)
# Later: retrieve exact dataset version used for any run
artifact = run.use_artifact("training-dataset:v3")
artifact.download()
**W&B Tables**:
- Log tabular data, images, audio, video, and text as interactive tables
- Compare model predictions across runs — see which examples improved or regressed
- Great for NLP: log input text, expected output, model output side-by-side
**W&B vs MLflow vs Neptune**
| Feature | W&B | MLflow | Neptune |
|---------|-----|--------|---------|
| UI Quality | Excellent | Good | Good |
| Sweeps/HPO | Built-in | External | Basic |
| Self-hosting | Yes (paid) | Yes (free) | Yes (paid) |
| HF Integration | Excellent | Good | Good |
| Collaboration | Excellent | Limited | Good |
| Free Tier | Generous | N/A (self-host) | Limited |
Weights & Biases is **the experiment tracking platform that turned ML research into a collaborative, visual, and reproducible practice** — by providing live training visualizations, automated hyperparameter search, and one-click experiment sharing with a SDK that integrates in three lines of code, W&B became the standard tool for ML teams who want to work faster and understand their models more deeply.