Training

The Training page launches LoRA SFT jobs and tracks their records.

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 training page.

Training config

The Training config card builds a single SFT job. Fields:

• Data version — pick an SFT version from the Data page. Click Upload SFT to upload a new JSONL inline.
• Training start (base model) — the frozen base to attach the LoRA adapter to, e.g. Qwen/Qwen3.6-35B-A3B.
• epochs, bs (batch size), lr (learning rate), rank (LoRA rank) — hyperparameters.

Click expand full command to preview the generated scripts/train.sh, then Start training to launch. The job runs as a background task.

Training only consumes self-built question-bank SFT

Training consumes SFT produced from your self-built question bank's passing trajectories. Official Benchmark records do not enter training — they are evaluation-only.

Training records

Training records lists every job: ID (version), Session, data version, row count, epochs/bs/lr/rank, the metric (NLL), timestamps, and actions (Details / Log / Delete). Click Details for the training curve and Log for the run log.

Training comparison

Training comparison lets you select two or more records and overlay their metrics — useful for seeing whether a new data version or hyperparameter set improved over the previous run. Select records with the checkboxes and click Compare.

Recent training artifacts

Recent training artifacts lists the most recent sampler_path and state_path outputs. The sampler_path is what you pass to inference; the state_path is what you pass to resume training.

Resume training

To continue from a previous version, point the training config at the prior version's state_path (see Checkpoints). Use --resume-state to restore weights, and add --resume-with-optimizer to also restore the Adam optimizer state for a true continuation.