Quickstart¶
Train an open model on your own data, then deploy it as a callable service.
This guide takes you from an SFT dataset to a trained model and its deployment. You need a Mint Recursive account, training examples, and an API key for inference. Review Service Pricing before submitting work.
1. Prepare your data¶
Upload a JSONL file in My data (我的数据). Each line is a JSON object containing a messages conversation and the desired response:
{"messages":[{"role":"user","content":"Classify this feedback: delivery was quick."},{"role":"assistant","content":"Delivery: positive"}]}
{"messages":[{"role":"user","content":"Classify this feedback: the packaging was damaged."},{"role":"assistant","content":"Packaging: negative"}]}
These examples illustrate the format. Prepare a representative dataset for an actual training run. After confirming the upload, wait for validation and select a version with valid structure and a nonzero sample count. See Data.
2. Train a model¶
Open Training (训练) and choose a base model and your dataset version. For the first run, start with the default SFT configuration and give the job a recognizable name.
Review the summary, run the training precheck, resolve any blockers, then submit. Serverless resources are scheduled by the platform; a job may queue before it starts.
Track the job in My training (我的训练). After success, find the resulting version in My models (我的模型). A successful job produces an artifact; evaluation determines whether it is suitable for your task.
3. Deploy your model¶
Deploy the selected version from My models, or select it on the Deploy (部署) page. Choose an available deployment method and submit. Confirm that loading, health, and routing are ready before making requests.
A model version and its serving deployment are separate resources. See Deployment for methods and readiness checks.
4. Make your first request¶
Open Model calls → API access (API 接入) and choose your model and an existing API key. Use the Base URL, Model ID, and request example shown there. Model version IDs, checkpoint paths, and API Model IDs are not interchangeable.
Verify a response with a representative prompt, then use Evaluation to compare the base and trained models. Keep evaluation examples separate from training data.