Skip to content

Mint Recursive

Mint Recursive's web console runs a recursive self-improvement loop: the model solves problems, the trajectories that pass become SFT training data, the trained model solves more problems, and the cycle repeats — so the model keeps improving without hand-written training data.

All compute runs on remote Mint Recursive servers. You need no local GPU — only a browser, a Mint Recursive API key, and (for SDK use) a Python environment.

Mint Recursive UI is Chinese-only

The Mint Recursive web console is currently localised in Chinese only. The English documentation below describes the pages and actions in English; the in-product labels you see will be Chinese. Screenshots are therefore omitted from this English documentation and provided only in the Chinese version.

Accessing Mint Recursive

Open the training workspace through Playground in the console. Sign in at mint.macaron.im (outside mainland China) or mintcn.macaron.xin (mainland China), then click Playground in the sidebar to enter the workspace through single sign-on. You can also open mindforge.macaron.xin directly.

The training loop

solve → score → train → solve again (with a stronger model) → ...
  1. Solve — the model reads each problem and generates solution code.
  2. Score — the code is executed and compared against expected results; each attempt is marked pass/fail.
  3. Training data — only passing trajectories are kept and turned into SFT data.
  4. Train — Mint Recursive LoRA SFT fine-tunes the model, producing a new version.
  5. Loop — the new version keeps solving, producing more correct trajectories.

You maintain the harness (the scoring and training pipeline); you do not hand-write training data. The model produces its own training data and keeps improving.

Platform features

The web console gives every operation in the loop a visual surface:

Page Function
Overview Resource status, Mint Recursive connection, running tasks, model-version evolution, evaluation-score comparison
Training Launch LoRA SFT training, pick base model and hyperparameters, resume from a checkpoint, view training curves
Evaluation Launch official SciCode Benchmark or question-bank evaluations, view details, heatmaps, pass-rate comparisons
Question Bank Browse question-bank versions, preview problems, launch question-bank eval, generate SFT data from passing trajectories
Data Manage SFT training data and question-bank data, upload/download files, preview rows
Guide Concept reference: checkpoints, using a trained model, switching the Mint Recursive backend

All compute-heavy operations (evaluation, training, data generation) run as background tasks; you can watch progress and logs live in the sidebar task panel.

Next steps