Skip to content

Question Bank

The Question Bank page manages your self-built problem sets and is the source of SFT training data: each problem carries its own test_data.h5 for scoring, and the model's passing trajectories can be turned into SFT.

Chinese-only UI

MindForge's web UI is localised in Chinese only. The field 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 question-bank page.

Question-bank evaluation

The Question-bank eval card evaluates a model version against a question bank:

  • Question-bank version — pick an uploaded bank. If none is available, the card shows an empty state (upload a question ZIP on the Data page first; it generates data/banks/<version>).
  • Model version — the version to evaluate.
  • Temperature, Concurrency.
  • Use background prompt, Enable thinking.

Click expand full command to preview scripts/bank_eval.sh, then Start question-bank eval.

Self-built banks generate SFT

Each problem in a self-built bank is scored by its own test_data.h5. Trajectories the model gets right can be turned into SFT training data — this is the loop's training-data source. Official benchmark results are evaluation-only and never become training data.

Problem preview

Problem preview lets you inspect a bank's problems before running an evaluation.

Question-bank evaluation records

Question-bank eval records lists past question-bank runs with pass rates and throughputs. From a record you can generate SFT data from the passing trajectories (the "solve → SFT" action on the Data page's text-bank table also kicks this off).

Adding a question bank

Upload a question-bank ZIP on the Data page (Question data or Text question bank section). After upload, the bank appears here as a selectable version.