Data

The Data page manages everything MindForge trains and evaluates against: SFT training samples, question-data ZIPs, and text question banks.

Chinese-only UI

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

Training data

Training data lists SFT JSONL files. Columns: Data version, Name, Filename, Trajectories, Problems, Token, Provenance (e.g. "→ training v3" when consumed by a training run, or "unused" when unused), and Actions (Details / Download / Delete).

• Upload training sample — upload a new SFT JSONL.
• Validate training samples — sanity-check format and counts before training.

Recommended source: passing trajectories

The recommended SFT source is passing trajectories from a question-bank evaluation (see Question Bank). You can also upload your own JSONL for a first run.

Question data

Question data manages question-data ZIPs — the binary bundles that include each problem's test_data.h5 scorer. Upload question ZIP uploads a new bank; Validate question data validates it. Uploaded banks appear as selectable versions on the Question Bank page (under data/banks/<version>).

Text question bank

Text question bank lists text-format question banks. Columns: Bank, recipe, Problem count, Path, Actions. The solve → SFT action runs the model on the bank and turns passing trajectories into SFT training data, which then appears in the training-data table above.

• Upload text question-bank ZIP — upload a new text bank.
• Validate format — sanity-check the bank's structure.