PRD — Live-streaming training-run dashboard (browser viewer)

Goal

Ship a browser page (/train-viewer.astro) that drag-drops a tinygpt train run-history directory (or watches it live via OPFS on the same machine) and renders the canonical live-training charts: loss + grad-norm + LR + tok/s over steps. Drag-drop pattern mirrors eval-leaderboard.astro and sae-timeline.astro — no server, the browser parses the on-disk artifacts.

Removes the “guess what’s happening at hour 6 of a run” tax. Today tinygpt train prints a line per step to stdout; you scroll the terminal hoping you saw the right thing.

Why now

• Every other interpretability surface has a viewer page now (eval-leaderboard, sae-timeline). Training-run is the one with no visual companion, and it’s the one users stare at the longest.
• The shipped train hook already writes per-step JSONL (loss, lr, grad-norm, tok/s) — there’s a stream to plot. The viewer reads what exists; no instrumentation change.
• Building this on the existing Astro + Chart.js-or-similar stack is ~1 day; building a W&B integration would be 3–5 days of auth + per-run-key wrangling for a project that explicitly wants self-contained tooling.

Scope — in

• web/src/pages/train-viewer.astro — drag-drop a directory (Chrome’s webkitdirectory); browser-side ingest of the run-history JSONL files.
• Charts (single page, vertical stack):
  • Loss vs step (train + val on the same plot; val is dotted)
  • Grad norm vs step (log scale)
  • LR vs step
  • Tok/s vs step (steady-state throughput)
• Optional sidecar: if the run was started with --eval-every N (E8 hook), the eval-leaderboard JSONL is read and overlaid as star-markers on the loss chart at the eval steps.
• Live mode (same-machine only): if the path is under /runs/<name>/history.jsonl, set an OPFS file-handle and poll every 5s. Works in Chromium-class browsers; documented as a Chrome- only feature.
• A “what to look for” panel below the charts: short tooltip text (“grad-norm > 3× rolling mean = spike risk; see B12”) + links to the relevant docs.

Scope — out

• W&B / TensorBoard integration. Out of scope by design — this is self-contained tooling.
• Multi-run overlay (compare two runs on the same chart). Useful but adds complexity; defer to V2.
• Mobile-responsive layout. Desktop-only — you’re watching a 4- hour run, you’re at a desk.

Files to touch

Don’t touch

• tinygpt train itself — viewer reads what it already writes.

Acceptance criteria

• Page loads at /train-viewer.astro in dev (pnpm dev) and prod (pnpm build && pnpm preview).
• Drag-drop a recent runs/<name>/ directory; loss / grad-norm / LR / tok/s charts render within 500ms.
• If an eval JSONL is present, star-markers overlay on loss at every eval step.
• Live mode: edit history.jsonl with a manual append; charts update within 5 seconds.
• Page is reachable via the existing /docs/[slug] route + linked from the training guide.

Reference patterns

• web/src/pages/eval-leaderboard.astro — exact template for drag-drop ingest + render. Steal the dropzone + the chart wrapper.
• web/src/pages/sae-timeline.astro — secondary template (time-series rendering).
• tinygpt train’s existing history.jsonl format — already documented in train-controls-gap-closure.md. No new format.

Open questions

• Chart library. Recommendation: whatever eval-leaderboard.astro already uses (don’t add a new dependency). If it’s Chart.js, stay on Chart.js.
• Live-mode polling vs file-system observer API. Recommendation: 5s polling. The observer API is Chrome-only AND requires permission prompts; polling is simpler and works on first load.