The web UI

mlp ui --root <store>        # serves on 127.0.0.1:7327

Localhost-bound by design; from another machine, tunnel like TensorBoard: ssh -L 7327:localhost:7327 <box>. Read-only except append-only annotations. On a served store with users configured (Deployment), the UI shows a sign-in page first; annotations are then authored as the signed-in user (writer role required), and the top bar carries your username and a sign-out link. Pages: Experiments (landing, run aggregates) → experiment (runs table, boards, experiment artifacts) → run (tabs: Overview | Metrics | Artifacts | Code), plus Boards (the gallery of all agent-authored boards — see Boards), Graph (lineage DAG), Search, and Diff.

Open runs show a liveness chip driven by the client’s heartbeat: a pulsing open · live while the training process is breathing, open · stale (hover for the last-seen time) when it stopped, plain open for runs with no liveness data.

The metrics dashboard

The run page’s Metrics tab is a configurable dashboard, stateful per run: layout, series selection, chart types, toggles, and zoom are stored in your browser (localStorage) and restored whenever you come back to the run.

  • Default layout: one panel per logged metric, in a 1/2/3-column grid (columns switcher in the top bar). New metric names appearing mid-run get a panel automatically until you start curating the layout — after that they’re listed as unplotted instead of disturbing it. reset layout regenerates the default.

  • Stat strip: big-number tiles above the grid with the latest value per metric (plus a step tile and a subtle trend arrow), updating live — at-a-glance training status. Toggle with stats.

  • Panels: drag the ⠿ handle to reorder (drop target highlights), ⛶ for full width, ✕ to remove, editable title, and a footer with last / min / max per series.

  • Panel editor (⚙): choose any set of series for the panel; chart type line / scatter / bar; x-axis = step, wall time, elapsed, or another metric (series joined on step — plot anything vs anything, e.g. val_loss vs lr); log-y; EMA smoothing; and exact numeric min/max for both axes.

  • Zoom: drag a rectangle on any plot (the selection is visualized) to zoom both axes; double-click or ⟲ restores. Zoom is per-panel. On a live run, a zoomed panel’s axes are pinned — streaming data keeps arriving but never resets your view; un-zoomed panels keep following new data.

  • Crosshair sync: hovering one panel shows the cursor on every panel with the same x-axis, with per-panel value readouts.

  • Export: per panel, png (the rendered plot) and csv (the panel’s series, always full resolution). Left-click downloads; right-click copies to the clipboard (the image itself for png, the values as text for csv) — paste straight into a chat, doc, or sheet.

Decimation (automatic)

Series above ~4k visible points render as a min/max envelope (~2k buckets, ≈2 points per pixel): every bucket keeps its minimum and maximum, so rare spikes always survive — a fixed-rate subsample would silently drop them. It re-resolves on zoom: decimation runs over the visible x-range, so zooming into any region recovers full detail automatically. There is deliberately no rate knob — the correct bucket count is a function of plot width, not preference. Stats and CSV export always use the full-resolution data; a footer note (“envelope of N pts”) marks when a panel is rendering the envelope.

The artifacts tab

Finder-style view controls on two levels, both remembered:

  • Page-wide: ▤ groups (default — artifacts grouped by media type: images, audio, video, tensors, checkpoints, text, other; grouping uses the logged media_type, falling back to the file extension) or ☰ list (one flat table).

  • Per group: ⊞ grid (default — content-first tiles: image groups show actual thumbnails, videos their first frame, everything else an icon tile) or ☰ detail rows.

Clicking an artifact — tile or row — opens an inspector pane on the right (the gallery stays visible beside it; it stacks below on narrow windows):

  • Images render inline; audio and video get players.

  • Text/CSV/JSON files show a (truncated) text preview.

  • Tensors (.npy, and .npz member-by-member) open the tensor viewer: slide through the leading axes with per-axis faders (the plot scrubs live) and view the remaining axes as a line plot (last axis) or heatmap (last two, viridis-colored). The normed checkbox (default on) fixes the color scale / y-axis to the whole tensor’s min/max so the range stays constant while sliding; off, each slice auto-scales. The whole-tensor range is scanned exactly below 32 MiB and estimated from an even sample above (marked ≈sampled). Reading is pure-Python on the server (no numpy/torch dependency), seeks directly to the requested slice, and strides large slices down to display resolution — a multi-GB array is never loaded whole. Plain numeric C-order arrays only; .npz members larger than 32 MiB can’t be sliced (zip entries aren’t seekable).

  • Checkpoints (.pt, .ckpt, .safetensors) are download-only: loading them requires the training framework (and unpickling untrusted checkpoints server-side would be unsafe). Prefer logging inspection tensors as .npy/.npz alongside the checkpoint.

Live streaming

Open runs stream metrics over SSE (/api/runs/<id>/metrics/stream); the dashboard batches updates (250 ms) and the stream ends itself when the run leaves open.