Skip to content

Troubleshooting

Common MindForge issues and how to resolve them.

Chinese-only UI

MindForge's web UI is localised in Chinese only. The error states described below appear as Chinese strings in the product; this page describes them in English. For labelled screenshots see the Chinese troubleshooting page.

Resource monitoring shows a connection error

The Resource monitoring card shows a red Service tile and dashes for all metrics.

  • Cause: the MindForge backend cannot reach the configured MinT server, or the server's get_server_capabilities / resource_status surfaces are unreachable.
  • Fix: open Run settings and verify the MinT URL and MinT Key. Clear the URL field to restore the server default, or point it at the correct endpoint (mintcn.macaron.xin for Mainland China, mint.macaron.im otherwise). Click Refresh after saving.

All metrics show -

  • Cause: the 5-minute perf_recent window has no data yet (no recent training/inference activity), or the backend returned no values for the requested fields.
  • Fix: run any training or inference job, then click Refresh. Missing fields display - by design; they do not indicate a service-health problem.

Question-bank page shows the empty state

  • Cause: no question bank has been uploaded yet.
  • Fix: upload a question-data ZIP or a text question-bank ZIP on the Data page. After upload, the bank appears as a selectable version under data/banks/<version>.

Training fails immediately

  • Cause: the SFT data version is missing or empty, the base model is unavailable on the configured backend, or the hyperparameters are out of range.
  • Fix: in the Training page's Training records, click Log on the failed row for the error. Verify the data version on the Data page (run Validate training samples), confirm the base model is served by your backend, and reset hyperparameters to defaults (epochs=3, bs=4, lr=2e-4, rank=16).

Playground entry shows "coming soon"

  • Cause: the dashboard's Playground link is gated in some sessions.
  • Fix: open mindforge.macaron.xin directly β€” the TokenFactory SSO still logs you in. See Quickstart access notes.

Inference returns "model not found"

  • Cause: the sampler_path is not yet indexed on the MinT server, or the wrong base URL / suffix is used.
  • Fix: run client.models.list() to confirm the sampler_path is in the server's index. Ensure the OpenAI-compatible base_url includes the /oai/api/v1 suffix (the mint SDK's MINT_BASE_URL does not).