Your datasets, ready for training
Import CSV and JSONL files into a shared workspace. Review schemas, identify invalid examples, and reserve a holdout set before training.
A single platform for dataset preparation, GPU accelerated fine-tuning, evaluation, and inference. Keep the context behind every model version, from its source data to its API endpoint.
Product preview. Capabilities described below are the target early access scope; availability will be confirmed during onboarding.
Import CSV and JSONL files into a shared workspace. Review schemas, identify invalid examples, and reserve a holdout set before training.
Choose from a planned catalog of Llama, Mistral, Qwen, and Gemma models. Select a foundation that fits your task, model license, and compute budget.
Configure low rank adapters and quantized training with QLoRA. Set hyperparameters, save configurations, and compare experiments in one place.
Track loss curves, epochs, GPU utilization, and job status. Review evaluation results alongside the base model before choosing a checkpoint.
The deployment workflow turns an evaluated model into an inference endpoint. Keep model versions connected to the data and jobs that created them.
Integrate with a familiar chat completions pattern. Keep model calls separate from application logic so your team can iterate on either independently.
Start with an open model that fits your task. The initial catalog is planned around these model families; exact versions and configurations will be confirmed during early access.
| Model family | Adaptation workflow | Bring your own data | Availability |
|---|---|---|---|
| Meta Llama | LoRA / QLoRA | CSV and JSONL | Planned catalog |
| Mistral | LoRA / QLoRA | CSV and JSONL | Planned catalog |
| Qwen | LoRA / QLoRA | CSV and JSONL | Planned catalog |
| Gemma | LoRA / QLoRA | CSV and JSONL | Planned catalog |
Model availability depends on architecture compatibility and license terms. Adapter and model usage remain subject to each foundation model’s license.
Training and inference are at the heart of the product architecture. Here is how we plan to use the accelerated computing stack.
GPU accelerated tensor operations for model training and serving.
Optimized inference execution for supported model architectures.
Containerized model serving as a future deployment option.
Training and evaluation tooling for reproducible model development.
Technology roadmap, not a claim of certification, partnership, or current integration. No NVIDIA affiliation is implied.
Bring a use case. Explore a model workflow built for your team.
Request access