Support

How long does training take?

Most fine-tuning jobs complete in 30–90 minutes depending on dataset size and the base model. You'll receive a notification when done.

What base models are supported?

We currently support Llama 3.2 (1B, 3B, 8B), Mistral 7B, and Phi-3 Mini. More models are added regularly.

Is my data private?

Yes. Your datasets and model adapters are stored in private AWS S3 buckets. We never use your data to train our own models. See our Privacy Policy for full details.

Can I export my fine-tuned model?

Yes. You can download the LoRA adapter weights from the Models page at any time. The adapter works with any vLLM or Hugging Face PEFT-compatible setup.

How is usage billed?

Subscription plans are billed monthly. GPU training time is metered separately and billed at the end of each cycle based on actual compute used (~₹12/hr for spot GPU).

How do I cancel my subscription?

Go to Dashboard → Billing → Cancel subscription. You'll stay on your current plan until the end of the billing cycle. No hidden fees.

My training job failed — what do I do?

Open the job in Dashboard → Jobs and check the error message. Common causes: dataset formatting issues (must be JSONL), or dataset too small (<20 examples). Email us with the Job ID and we'll investigate.

Do you offer refunds?

Yes — full refund within 7 days of your first paid subscription. After that, refunds are issued for verified service outages. See our Terms of Service.