1.What counts as 'delivered'
An order is considered delivered when we have sent you a signed download link for the trained model weights and the evaluation report associated with the tier you ordered. The signed download URL is active for at least 14 days (Starter / Studio) or 30 days (Atelier) from the delivery email; once you click it and download, or the window passes, the artifact is considered transferred. After the window closes we delete our copy of the weights — they live with you, not us. We do not host the model on your behalf and we do not run inference for you.
Once an order is delivered, the work product is in your possession and the order is non-refundable.
2.Cancellation before training begins
You may cancel your order at any time before the training job is queued for execution. To cancel, email us at 0xcircuitbreaker@protonmail.com with your order ID.
- Before payment is captured: nothing to refund. The order is voided.
- Payment captured but training not yet started: 100% refund, processed within 5 business days via the original payment method.
3.Cancellation during training
Once a training job is in progress, GPU compute is being consumed in real time. If you cancel mid-job:
- Cancellation request received in the first 25% of the training run: 50% refund (we eat the partial-job compute cost).
- Cancellation request received after 25% progress: No refund — by that point most of the compute has been spent.
Atelier-tier orders include one revision round, so if you're unhappy with the initial training pass, we'd rather adjust hyperparameters and re-run than refund.
4.Operator failures (our fault)
If a job fails due to our infrastructure (training crashes we can't recover, severe quality regressions traceable to operator error, our inability to deliver within a reasonable extension of the stated turnaround window), you get either of the following at your choice:
- Full refund within 5 business days, or
- Free re-run on the same tier, with one round of hyperparameter adjustment included.
5.What is not eligible for a refund
- Successful delivery where the customer is dissatisfied with the model's accuracy or behavior. Fine-tuning outcomes depend heavily on the quality and fit of the training data you provide. We cannot guarantee a specific accuracy level on data we did not curate. Tier-included evaluation reports describe the resulting model honestly so you can decide whether to deploy.
- Customer-supplied training data that violates third-party rights (copyright, PII without consent, etc.). Our terms of service require lawful, properly-licensed data. Orders cancelled because we discover non-compliant data are non-refundable.
- Crypto orders (NOWPayments) where exchange-rate movement between order time and refund time creates a USD/crypto delta. We refund the originally captured USD amount; we do not refund the crypto-denominated amount that may have appreciated.
- Subjective dissatisfaction with the chosen base model's general capabilities. If you ordered a fine-tune on Llama 3.1 8B and discovered after delivery that the base model isn't well-suited to your task, that is not a refund event — the wizard described each model's strengths before you ordered.
6.How to request a refund
Email 0xcircuitbreaker@protonmail.com with:
- Your order ID (format:
LGS-XXXXXXXX-XXXXXX) - The reason you're requesting a refund
- The email address associated with the order
We respond within one business day. Approved refunds are processed within 5 business days via the original payment method.
7.Disputes and chargebacks
Before initiating a chargeback through your bank or card issuer, please email us first. Most disputes are resolved faster directly. Chargebacks initiated without a prior good-faith refund request will be contested with documentation of the order and delivery.
8.Changes to this policy
DDG may update this refund policy from time to time. The version in effect at the time of your order applies to that order. Material changes will be reflected in the "Last updated" date at the top of this page.