Data

Manage the examples your model learns from, with a traceable dataset version for every run.

Mint Recursive connects uploaded files, validation results, and training jobs through versioned datasets. Manage them in My data (我的数据).

SFT data format

The current upload flow accepts JSONL with one JSON object per line. Use messages to represent a conversation: user messages describe the task and assistant messages provide the desired response.

{"messages":[{"role":"user","content":"Classify: support replied quickly."},{"role":"assistant","content":"Support: positive"}]}

This illustrates structure only. Build examples around real tasks with consistent labeling and output formats. Review model-generated responses before treating them as correct answers.

From upload to a trainable version

Choose Upload dataset (上传数据集) in My data, name the dataset, and select a JSONL file. The interface also offers pasted JSONL input. Confirm the upload to begin server-side validation.

Dataset versions and validation results · Screenshot placeholder · final product capture to follow

Check validation state and valid sample count before training. A row appearing in the list does not itself mean upload or validation succeeded.

State	Action
Structure valid and nonzero valid samples	Select the version for training and run the precheck
Structural errors	Inspect the error, correct the data, and upload again
Upload incomplete / upload_expired	Upload again and wait for confirmation
Zero valid samples	Check file content and structure before training

Keep data traceable

Name datasets by task and revision. Record the exact version used by each training run; upload a new version when examples change.

Structural validation does not verify answer correctness or dataset coverage. Review quality and reserve separate evaluation data from the outset.

Next: Training · Question Bank