Skip to main content
Telescope UI runs locally on your machine and pulls training data from Weights & Biases into a local database for fast querying and visualization. 
Trainer → W&B → Telescope UI → DuckDB → Dashboard

​
Run discovery

When you launch the UI and provide your W&B credentials, you add the W&B projects you want to monitor. Telescope polls these known projects for runs tagged with telescope and lists any matches in the sidebar. You can hide a run from the UI by adding the telescope-ignore tag in W&B.

​
Data sync

During training, the trainer continuously uploads structured data (metrics, timeline events, rollouts, evals, GPU stats) to W&B as compressed archives. The UI downloads these archives, extracts the parquet files inside, and inserts them into a local DuckDB database. The sync is incremental — the UI tracks what it has already ingested and only downloads new data. For a running training, it polls every few seconds so the dashboard stays up to date in near real time. There are three categories of data:
  • Events — timeline events, system metrics, and logs (orchestrator events, trainer events, inference timings, GPU/CPU/vLLM metrics, trainer/inference process logs). Uploaded as rolling 30-minute blocks.
  • Steps — per-step rollout data (prompts, completions, rewards, eval results, golden answers). Uploaded as blocks of up to 500 steps.
Both categories also have a tail archive that contains the most recent data, updated every 5 seconds, so the UI can show fresh results without waiting for a full block to finalize.

​
Querying

Once data is in DuckDB, the dashboard frontend queries it through the local FastAPI server. All pages — Overview, Metrics, Rollouts, Rollouts Discarded, Timeline, Infra, Evals — read from this same local database, which makes queries fast regardless of how much data your training run has produced.