Skip to content

Quickstart

This page walks through accessing MindForge and running one turn of the recursive self-improvement loop: upload SFT data → configure training → start → inspect the record → use the resulting sampler.

Chinese-only UI

MindForge's web UI is localised in Chinese only. The button and field names below are given in English; the labels you see in the product will be Chinese. For labelled screenshots see the Chinese quickstart.

Access MindForge

MindForge lives at mindforge.macaron.xin and shares its account with the MinT dashboard. Two ways in:

  1. From the dashboard — sign in at https://mintcn.macaron.xin (Mainland China) or https://mint.macaron.im (outside), then click Playground in the sidebar. You are single-sign-on'd into MindForge.
  2. Direct URL — open https://mindforge.macaron.xin. If not already signed in, you are redirected through the dashboard's TokenFactory SSO; log in there and you bounce back into MindForge.

If the dashboard's Playground entry shows "coming soon", open mindforge.macaron.xin directly — the SSO still works.

Step 1 — Upload SFT training data

Open the Data page and click Upload training sample. Pick a JSONL file of prompt → response pairs. After upload it appears in the training-data table with a version id like sft_v1. The recommended source is passing trajectories from a question-bank evaluation (see Question Bank), but any SFT JSONL works for a first run.

Step 2 — Configure training

Open the Training page. In the Training config card:

  • Data version — select the SFT version you just uploaded.
  • Training start (base model) — pick a base, e.g. Qwen/Qwen3.6-35B-A3B.
  • Hyperparameters — epochs, bs (batch size), lr (learning rate), rank (LoRA rank). Defaults like epochs=3, bs=4, lr=2e-4, rank=16 work for a first run.

Click expand full command to preview the generated scripts/train.sh before running.

Step 3 — Start training

Click Start training. The job runs as a background task; watch it in the sidebar task panel (bottom-left) or the Training records table. When it finishes, the record shows the metric (NLL), duration, and a session id.

Step 4 — Inspect the record

In Training records, click Details to see the training curve, or Log for the run log. The record's version id (e.g. v6) is what you will infer against.

Step 5 — Use the trained model

Training produces a sampler_path (a tinker://.../sampler_weights/... reference). Use it for inference in any of the three ways documented in Using a trained model — the simplest is the OpenAI-compatible API with model set to the sampler_path.

Next

  • Training — full reference for the training panel.
  • Checkpoints — State vs. Sampler checkpoints.
  • Run settings — point MindForge at a different MinT backend.