Learn your labels
Fine-tune a small pretrained model on the examples you already have. Your categories define the task.
Small models.
Made for your work.
Bring your examples. Describe the task. SotaYard handles the training, evaluation, and deployment of a model that learns your labels.
Train your first modelWhy SotaYard
Your examples.
Your decisions.
Your model.
Route support tickets, categorize records, or tag incoming messages. Teach a small model the distinctions that matter to your work.
Measure before you ship
See results on held-out examples, compare a baseline, and inspect errors before a model can go live.
Put it to work
Deploy a qualifying version to an authenticated API. Try it in the playground and flag uncertain results for review.
Keep improving
Save corrections, export new examples, and train another version. Roll back to a previous deployment when you need to.
From data to deployment
Bring the task.
We handle the training.
One workspace for your data, training runs, evaluation reports, and deployed models. Every version has a result you can inspect.
Support routingIllustration
01 / Your dataset
A few examples.
A clear task.
support-tickets.csvInput: message · Answer: label
CSV| Message | Label |
|---|---|
| Please resend my invoice | billing |
| The export keeps failing | technical |
| I was charged twice | billing |
You supply the answers. The model learns the pattern.
Support routingIllustration
02 / Managed training
One task.
A model of its own.
- 01Check your examplesLabels, duplicates, and class balance
- 02Fine-tune the modelPretrained encoder → your labels
- 03Evaluate a candidateReserved examples, independent of training
Follow each run from preparation through evaluation.
Support routingIllustration
03 / Evaluation
Evidence comes
before deployment.
TrainSelectCalibrateTest
- Accuracy
- How often accepted answers are right
- Coverage
- How much work the model can accept
- Baseline
- A reference to compare the model against
Only a model that qualifies can be deployed.
Support routingIllustration
04 / Your model, available by API
Ready for the
next real example.
POST
/api/v1/projects/:id/predict{
"inputs": {
"message": "Please send my invoice"
}
}Example labelbilling
Test in the playground. Connect with an API key.
What can I build with SotaYard?
The current beta trains text and structured-record classifiers: one label per example, with 2–12 possible labels. Use it for tasks such as ticket routing, intent detection, and record categorization. Generative chat, image models, and arbitrary training code are not supported in this beta.
What data do I need?
Bring a CSV with a header row or JSONL with one flat object per line. Include input fields and the correct label. The beta supports up to 2,000 rows and 5 MB per file. You need at least 80 distinct inputs and 12 examples per class; several hundred diverse examples usually give a more useful evaluation. Related examples should share a group ID.
Do I need to write training code?
No. The workspace handles data preparation, fine-tuning, evaluation, deployment, and test predictions. Describe the task, map your columns, and set your quality requirements. Connecting the deployed API to your application uses a standard authenticated HTTP request.
Will every trained model be deployed?
No. A completed run must qualify against your accuracy and coverage requirements and its baseline before it can deploy. Evaluation uses reserved examples and conservative confidence bounds. If the evidence is insufficient, the report explains why and you can add genuinely new examples.
What does the beta include?
The beta is free and bounded: up to 3 training runs per account per day, 1 active run at a time, and 1,000 predictions per account per day. Each dataset can contain up to 2,000 rows and 12 classes. No credit card is required. See the guide for storage and project limits.
What happens to my data?
Your examples train models for your account, not models shared with other customers. You can delete a project and its datasets, model artifacts, and saved corrections. Usage records and one-way evaluation fingerprints remain to enforce limits and prevent test-data reuse. The privacy page explains processing and retention.
How do I improve a model after deployment?
Save corrected labels in the playground and export them as JSONL. Update your source dataset, add genuinely new examples, and train another version. Corrections do not silently change the model serving your API. You can deploy a qualifying candidate or roll back to the previous deployment.
Give your data
a job to do.
Start with your own labeled examples, or try the sample dataset. Train, evaluate, and serve a text classification model in one workspace.
Start building free