telescope-ui 0.1.21

Real-time visualization dashboard for LLM training runs

The visualization dashboard for Telescope — a framework for post-training LLMs with reinforcement learning, designed for reasoning and agentic capabilities.

Telescope UI

Telescope UI runs locally on your machine and syncs training data from Weights & Biases into a local DuckDB database, giving you a real-time dashboard to monitor runs, inspect rollouts, and analyze metrics.

Installation

pip install telescope-ui

Requires Python 3.11+.

Quick start

wandb login    # authenticate with W&B (once)
telescope      # starts the dashboard at http://localhost:8005

Telescope automatically discovers runs tagged with telescope in your W&B projects. The default training config already includes this tag, so runs appear in the UI automatically.

Flag	Description
--port PORT	Port to serve on (default: 8005)
--host HOST	Host to bind to (default: 127.0.0.1)
--no-browser	Don't open the browser automatically
--debug	Enable verbose logging

Documentation

Full documentation at docs.telescope.training.

Pages

Metrics

Rollouts

Timeline

Infra

Evals

Code Compare

Data storage

All synced data is stored locally at ~/.telescope/. Override with the TELESCOPE_DATA_DIR environment variable. Database compression is available through the sidebar menu and can reduce size by 20-30%.

License

MIT