Troubleshooting

    A top-of-funnel diagnostic checklist. Match your symptom to a section here, then click through to the canonical detailed page for the fix.

    Most problems fall into one of five categories: a dataset would not upload or train cleanly, a run failed or stuck mid-training, an exported GGUF will not load, an on-device app behaves badly, or Ertas itself is misbehaving. This page is the symptom-to-page redirect; the detailed fixes live in the section troubleshooting pages.

    Find your symptom

    SymptomWhere to go
    Dataset upload rejected; row counts off; columns missingDatasets troubleshooting
    Loss is not going down; training looks stuck; model output is emptyDatasets troubleshooting
    Run errored mid-training; "out of memory" messageHandling failures
    Run completed but artifacts are missingHandling failures
    GGUF will not load in Ollama or LM StudioVerifying exports
    Model loads but produces raw chat-template tokens in outputVerifying exports
    On-device inference is slow; runs out of memoryPerformance tips
    Model crashes the app after the device wakes from sleepShip: Android or Ship: iOS
    Studio canvas is blank or unresponsiveThis page, "Studio in the browser" below
    Login or billing issueThis page, "Account and billing" below

    If your symptom is not in this list, the FAQ catches the long tail of common questions.

    Studio in the browser

    These are the few symptoms that the section-specific pages do not cover.

    The canvas is blank or unresponsive

    Studio's canvas is built on standard browser primitives (SVG and 2D canvas); there is no WebGL or WebGPU requirement. When the canvas does not render, the cause is almost always an extension or a stale asset cache, not a graphics-API issue. Three things to check:

    1. Hard refresh with Ctrl+Shift+R (Cmd+Shift+R on Mac). Module updates occasionally need a fresh fetch of the static assets.
    2. Disable extensions that block or rewrite scripts on third-party sites (some privacy and ad-blocking extensions intercept the canvas-rendering code). The fastest test is an Incognito window with all extensions disabled.
    3. Try a different browser to isolate. If the canvas renders in Firefox but not Chrome (or vice versa), the difference is an extension or a browser-profile setting, not Studio.

    If all three pass and the canvas still does not load, file a bug; the support flow is in FAQ.

    Dataset uploads fail without an error

    A few causes:

    • File over your plan's dataset storage cap or row-count limit. Plan-specific caps live on the Pricing page; the Storage page explains how Data Craft tracks per-plan quotas.
    • JSONL not strict JSONL. One JSON object per line, no trailing comma, no surrounding array brackets. The JSONL format page covers the exact shape.

    Runs queue forever and never start

    When the GPU pool is busy at peak hours, queue time can extend. Two things to verify:

    • Your credit balance covers the run estimate. Ertas puts the estimated cost on hold when you press play; if your non-held balance is too low, the run is blocked at submission rather than queued. The Run panel surfaces this.
    • Your account is not subject to a billing hold. A billing failure (declined card, exhausted plan) blocks new runs until resolved. Resolve in the Billing settings.

    If both look fine and runs are still stuck, the queue is busy. Cancel and re-queue in an off-peak window.

    A model in the picker is locked

    The model requires a GPU tier above what your plan allows. Click the model card and Ertas surfaces an upgrade prompt. The Free-plan-eligible model list shows which models run on T4 (Free) versus A10G (paid).

    Account and billing

    SymptomFix
    Cannot sign inReset password from the login screen; if Google or GitHub OAuth fails, try the email-password flow as a fallback.
    Credits not refreshingFree-plan credits refresh on a fixed cadence; see Pricing for the current grant rate and the Credits and usage page for how refresh interacts with held credits.
    Billing failed but plan still says activeBilling failures grace for a short period before the plan downgrades. Update the card in Billing settings as soon as you see the warning.
    Upgrade did not take effectPlan changes take effect immediately but the previous plan's credits do not transfer. Reload after a successful upgrade to refresh the Run panel's plan tag.

    For anything billing-related not covered here, the support email in the footer is the path; logs and account state are server-side and a support agent can see them.

    Why these checks are quick

    The detailed troubleshooting in the section pages assumes you have already isolated the layer where the problem is happening: dataset, training, export, or on-device. This page is the layer-picker. If you skip the layer-pick and start debugging at the wrong layer, you waste time; if you pick the right layer first, the detailed page is short.

    What's next