Skip to content

Overview

The Overview page is the MindForge landing page. It shows the health of the MinT backend, what is running right now, and how your model versions compare on the SciCode benchmark.

Chinese-only UI

MindForge's web UI is localised in Chinese only. The tile and column names below are given in English; the labels you see in the product will be Chinese. For a labelled screenshot see the Chinese overview page.

Resource monitoring

The Resource monitoring card uses a fixed 5-minute window (perf_recent) and reports:

  • Service — the MinT service status (from get_server_capabilities).
  • Last 5-min requests — request count, split into training / inference.
  • Last 5-min training/inference p50 latency — training and inference p50 latency, with TTFT (time to first token) and save sampler sub-metrics.
  • Session status — session counts by state: active, idle, training, inference (from resource_status).
  • Recent active training throughput — training throughput (fb mean, step mean).
  • Recent active inference throughput — inference throughput (input, output).

Request counts and latency come from the groups surface. The rate card shows only active_compute_rate — the active compute rate. Performance-metric protocol mismatches do not change service health status; missing fields show -. Click Refresh to re-poll.

Tasks

The Tasks panel lists background jobs across four categories, each with a counter:

  • Official evaluations
  • Question-bank evaluations
  • Training
  • Data generation

Select a category to see its jobs with status, duration, and Details / Log links. The sidebar task pill (bottom-left) summarises the most recent completion.

SciCode evaluation comparison

SciCode evaluation comparison shows external reference models' SciCode scores for context. It only displays external reference models — not your local models or training versions. Local-model results live in the evaluation records and training curves. Scoring is SciCode single-generation sub-step pass rate.

Loop timeline

Loop timeline visualises the recursive self-improvement loop over time — when each solve → score → train cycle ran and which model version it produced.