ai.vita BETA

A human-approved training dataset.
Training examples are written and approved by the organisation's experts, so the dataset carries its knowledge and language.
An open-source model, tuned on your data.
Qwen 27B is fine-tuned (LoRA or full) on an on-demand GPU server, with progress, loss and cost monitored live.
Before and after, on the same questions.
The fine-tuned model is compared with the base model on a held-out test set, question by question.
A measured model that fits 20 GB.
The model is quantized to fit in 20 GB of memory and measured on the same questions before and after, so any quality loss is visible.
On-premise or in your cloud.
The finished model runs on a server with a single 20 GB GPU. You choose where: your own on-premise server, or a virtual server in your own account with the cloud of your choice. The training server is then shut down.

Every cost shown upfront.
The hourly price is shown before a GPU server starts, idle servers shut down automatically, and every action that spends money is confirmed with its amount.
Beta status: sign-in, sessions, users and activity work in the panel; training, servers and the other screens are currently previews with sample data. GPU provider selection is in progress. Voice, search and image modules are on the roadmap.
