Bread Wandb Viewer

Compare ML training runs side-by-side in VS Code - no browser switching, no waiting

Stop switching to your browser to compare training runs. Bread Wandb Viewer brings Weights & Biases visualization directly into VS Code. Compare multiple runs side-by-side with interactive charts, export AI context for Claude Code/Cursor/Codex, and analyze experiments without leaving your editor.

Perfect for machine learning engineers, deep learning researchers, and data scientists who want to stay in their development environment.

Multi-Run Comparison in Action

View multiple training runs in seconds - no browser needed

Compare W&B training metrics with interactive sidebar

Features

• 🔄 Multi-Run Comparison - Overlay multiple runs on the same charts with color coding and interactive toggles
• 🤖 AI Context Export - One-click markdown export for Claude Code, Cursor, and Codex with token counting
• 🎯 Interactive Charts - Zoom, pan, fullscreen, and smooth with EMA (exponential moving average)
• 📂 Automatic File Watching - Extension detects new .wandb files and refreshes automatically
• 📊 System Metrics - GPU utilization, memory, CPU, disk I/O tracking
• 🔍 Metadata Comparison - Side-by-side config and hyperparameter diff highlighting
• 🎨 Smart Grouping - Metrics auto-organized by prefix (loss/, train/, val/, gpu.0/)
• 🔒 100% Offline - Direct .wandb file parsing with protobuf, no API calls or internet needed
• 🎛️ Advanced Controls - Log scales, raw data overlay toggle, adjustable smoothing
• 📁 Folder Scanning - Automatically discover all runs in a directory

Quick Start

View a single run:

1. Click any .wandb file in VS Code Explorer
2. Charts and metrics appear instantly

Compare multiple runs:

1. Right-click any folder with W&B runs
2. Select "Bread Wandb Viewer"
3. Check/uncheck runs in sidebar to compare

That's it! No configuration needed.

Visual Demos

AI Context Generation

One-click AI context export for Claude Code, Cursor, and Codex

Features in Depth

🔄 Multi-Run Comparison

Compare training runs side-by-side to understand what hyperparameters and configurations work best - all without leaving VS Code.

• Overlay multiple runs on the same charts
• Color-coded run identification
• Interactive sidebar for toggling runs on/off
• Resizable sidebar for better workspace management
• Automatic folder scanning for all runs

Ideal for hyperparameter tuning, ablation studies, and experiment analysis. No need to switch to your browser to compare metrics.

🤖 AI Context Export

Export your training runs as AI-optimized markdown for coding assistants like Claude Code, Cursor, and Codex.

Perfect for:

• Debugging training code with AI assistance
• Analyzing hyperparameter impact on model performance
• Generating experiment summaries and insights
• Understanding why certain runs performed better

Features:

• One-click copy to clipboard or save to file
• Token count estimation for context planning
• Formatted comparison tables
• CSV metric data for analysis
• Configuration diffs highlighted

🎯 Interactive Charts

Advanced chart controls for detailed metric analysis.

• Smoothing: Adjustable EMA smoothing with real-time preview
• Zoom: Drag to select region and zoom in
• Pan: Shift+drag to navigate zoomed charts
• Fullscreen: Click expand icon on any chart
• Log Scales: Toggle X and Y axis logarithmic scales
• Raw Data Overlay: View smoothed and raw data simultaneously
• Auto-decimation: Large datasets (500+ points) automatically downsampled for performance

📂 Automatic File Watching

The extension automatically detects when new .wandb files are created or updated and refreshes the view without manual intervention.

• No refresh button needed
• Works seamlessly with running training scripts
• Detects new runs added to folders
• Updates existing run views when files change

📊 Metadata & System Metrics

View comprehensive run information beyond just training metrics.

Run Metadata:

• GPU type and count
• Python version
• CPU count and CUDA version
• Git remote and commit
• Runtime start timestamp

System Metrics:

• GPU utilization and memory
• CPU usage
• Disk I/O
• Memory consumption

🔒 100% Offline & Private

Your training data never leaves your machine.

• Reads .wandb files directly using protobuf
• No wandb CLI or API needed
• No internet connection required
• Works completely offline
• Zero telemetry on your training data (see Privacy section below)

Installation

Install from the VS Code Marketplace, or download the .vsix from releases:

code --install-extension wandb-viewer-0.2.2.vsix

Requirements

• VS Code 1.74.0 or higher
• Training runs created with wandb SDK 0.15+

Commands & Usage

Keyboard Shortcuts

Privacy & Analytics

This extension collects anonymous usage analytics to help improve the product. We take your privacy seriously.

What We Collect

• Feature usage (which features you use)
• Performance metrics (parse times, load times)
• Error events (crashes and bugs)
• Chart interactions (smoothing settings, zoom actions)

What We DO NOT Collect

• ❌ Run names or project names
• ❌ Metric values or training data
• ❌ File paths or code
• ❌ Hyperparameters or configurations
• ❌ Any personally identifiable information (PII)

How to Opt Out

Option 1: Via Settings UI

1. Open Settings (Cmd+, or Ctrl+,)
2. Search for "telemetry level"
3. Set Telemetry Level to "off"

Option 2: Via settings.json

1. Open Command Palette (Cmd+Shift+P / Ctrl+Shift+P)
2. Type "Preferences: Open User Settings (JSON)"
3. Add: "telemetry.telemetryLevel": "off"

The extension respects your editor's global telemetry.telemetryLevel setting. Learn more about VS Code telemetry.

Technical Details

For developers and power users interested in how this extension works.

Architecture

• Binary Parsing: Direct protobuf parsing of .wandb files (LevelDB-style format)
• No Dependencies: No wandb CLI, API, or internet connection required
• Performance: LRU cache (20 runs), LTTB decimation for large datasets, lazy chart initialization
• File Watching: Automatic detection of file changes with 1-second debouncing
• Chart Library: Chart.js 4.4.0 with zoom plugin for interactive visualizations

Supported Record Types

• HistoryRecord (per-step training metrics)
• ConfigRecord (hyperparameters)
• SummaryRecord (final summary stats)
• StatsRecord (system statistics)
• EnvironmentRecord (Python, GPU, host metadata)
• RunRecord (project name, run ID, display name)
• And 15+ other internal record types

Performance Optimizations

• Quick Metadata: Reads only first 16KB for fast folder scanning
• Metric Decimation: LTTB algorithm for datasets >500 points
• Lazy Loading: Charts initialized only when visible
• LRU Cache: Parsed run data cached (max 20 runs)
• Debounced Updates: File changes debounced to 1 second

Actively Maintained

• ✅ Current version: 0.2.2
• ✅ Compatible with VS Code 1.74+
• ✅ Tested with wandb SDK 0.15+
• ✅ Open Source - contributions welcome

Issues & Support

Found a bug or have a feature request?

• Report Issues: GitHub Issues
• Pull Requests: Contributions welcome!
• Documentation: Source Code

Why Bread Wandb Viewer?

Problem: Switching between VS Code and your browser to compare training runs breaks your flow and slows down iteration.

Solution: View and compare everything in VS Code. Multi-run comparison, AI context generation, and interactive charts - all without leaving your editor.

Result: Stay focused, iterate faster, and leverage AI coding assistants to analyze your experiments.

Made with ❤️ by Bread Technologies