LLM Studio: training custom models with result verification

LLM Studio is a Snaga platform tool for training your own language models for a specific domain. Unlike supervised fine-tuning on input/output pairs, training uses a domain verifier: the generated output is executed, and the score is computed from the actual execution result rather than from the text's similarity to an example.

How training works

A single training step:

The model generates output for a task.
The domain verifier executes that output and returns an objective score.
The score is used as a reward signal (outcome-based, RL approach).

This removes the limitation of imitation learning, where the model optimizes for text plausibility rather than for the correctness of the result.

Supported domains

Only domains with a deterministic verifier are active:

Code — the generated program is run against a test set; the score is the fraction of passing tests.
SVG — the generated vector graphic is rendered; the score is computed from the rasterized result.

Domains without a verifier are not enabled. The set expands as new verifiers are added.

Job configuration

Domain & persona — the target domain and the model's behavior profile.
Tier: Small or Medium — determines the GPU class and the amount of training.
GPU: SKU selection from the catalog; up to 2 droplets per job (parallel training).
Hyperparameters — set in the wizard before launch.

Workflow

create → training → export → served

A step-by-step wizard takes you from domain selection to launch.
A real-time metrics stream is available during training.
A job can be canceled or retried.

Output & integration

The model is registered as a provider and becomes selectable for any agent on the platform — a trained model immediately enters the working loop.
Weights are exported in GGUF format for local execution outside the platform.

Aspida: an open-source inference engine

Training and inference run through Aspida — an open-source inference engine the team develops as a separate technology to improve the safety of future communication between humans and AI. It is not a third-party dependency, but the platform's own work:

End-to-end encryption (E2EE) of training and inference; the server's public key is pinned — training data and model weights never leave the encrypted boundary.
Each job runs in an isolated, ephemeral GPU droplet that is destroyed on completion.
Open source: github.com/chabanov/aspida

Availability

The LLM Studio section in your workspace.

How training works

A single training step:

The model generates output for a task.

The domain verifier executes that output and returns an objective score.

The score is used as a reward signal (outcome-based, RL approach).

This removes the limitation of imitation learning, where the model optimizes for text plausibility rather than for the correctness of the result.

Supported domains

Only domains with a deterministic verifier are active:

Code — the generated program is run against a test set; the score is the fraction of passing tests.

SVG — the generated vector graphic is rendered; the score is computed from the rasterized result.

Domains without a verifier are not enabled. The set expands as new verifiers are added.

Aspida: an open-source inference engine

End-to-end encryption (E2EE) of training and inference; the server's public key is pinned — training data and model weights never leave the encrypted boundary.

Each job runs in an isolated, ephemeral GPU droplet that is destroyed on completion.

Open source: github.com/chabanov/aspida