Teach an open model your way of working.
Fine-tune on your reviewed data with SFT or DPO, using LoRA or 4-bit QLoRA. Tensorant manages the training workflow on GPUs in your connected cloud account.
Choose a frozen dataset, training objective, adapter method, and GPU.
Interactive illustration · sample data
A training run, with every stage in view. See how preparation, training, validation, and GPU cleanup fit together.
Your training choices. The infrastructure handled.
Control the settings that matter to your task, with readiness checks and a clear record of what ran.
Choose how the model learns
Use SFT to learn from expected answers or DPO to learn from chosen and rejected responses. Choose LoRA or QLoRA independently of the training objective.
Explore the featureStart with a small trial
Test your settings on a small sample before a full run. Review the trial report, recommendations, and estimate for the larger job.
Explore the featureUse your connected GPU provider
Train with RunPod, Verda, Lambda, AWS EC2, or Azure VMs. Check model fit and set runtime and hourly price limits before launching.
Explore the featureRun a complete experiment
Connect baseline evaluation, training, comparison, and cleanup in a persisted experiment. The workflow continues even when your client disconnects.
Explore the featureKeep the adapter. Release the compute.
Follow logs, validation metrics, and spend as a run progresses. Tensorant requests GPU cleanup after the run and surfaces recovery work when cleanup needs attention.
- Pinned inputs
- A frozen dataset and base-model revision make the run traceable.
- Baseline first
- Full training requires a completed baseline for the same version.
- Portable output
- Download the adapter or publish it to Hugging Face.
Build with Tensorant