vibetrack

Overview

Designed as a modern alternative to heavier tracking tools, Vibetrack stores experiment data using a thread-local SQLite WAL backend with bulk insertion and precache capabilities. It supports two primary API styles (TensorBoard-compatible SummaryWriter and module-level functions) and includes dynamic output viewers like Web UI, Gradio, console sparklines, Telegram, Slack, and an MCP server. Non-blocking error handling ensures that telemetry or notification failures never interrupt the core training loop.

Capabilities

• ▸TensorBoard-compatible logging API and module-level logging API
• ▸SQLite-backed storage with Write-Ahead Logging (WAL) and bulk flush optimization
• ▸Auto-discovered output viewers including Web, Gradio, Console, Telegram, Slack, and MCP
• ▸Cross-experiment comparison across scalars, hyperparameters, and summary tables
• ▸Optional system and GPU metrics collection
• ▸Best-effort, non-blocking network error handling for external notifications

Best for

Tracking scalars, images, audio, video, text, histograms, and embeddings during machine learning model training., Comparing multiple training runs side-by-side using a responsive web dashboard or CLI tools., Real-time experiment monitoring and notifications via Telegram, Slack, or terminal output., Ingesting remote training logs over HTTP with optional bearer token authentication.

Works with

Security considerations

Exposes an optional HTTP ingest server and Model Context Protocol (MCP) server. Ensure proper bearer tokens and network restrictions are applied when hosting publicly to prevent unauthorized log ingestion or data leakage.

Categories