Everything people ask before they train their first model.
If something is missing, ask on the contact page and we will add it here.
Getting started
What you need before your first run.
Do I need to know PyTorch or CUDA?
No. You describe the task and the output you want. The agent picks the base model, writes the trainer config, calculates the memory budget and runs the job. Every generated file stays visible if you do want to read it.
What do I need to bring?
Some idea of the task, and a labelled dataset if you already have one. Compute runs on Tensovy managed infrastructure, so there is nothing to set up, or on your own cloud account if you would rather connect a key. If you have no dataset yet, you can search Hugging Face and Kaggle from inside the workspace and import a starting one.
How long does a first run take?
A small adapter fine tune on a single consumer class GPU is typically under an hour end to end, including data preparation and weight export. Larger models and datasets take longer.
Can I start small?
Yes, and most people should. A first adapter fine tune runs on a single consumer class GPU, so the compute involved in finding out whether a custom model works for you is small. Plan pricing is not published yet, so tell us the workload and we will quote it.
Compute and billing
How each compute route is billed in practice.
Who pays for the GPU time?
That depends on the route. On Tensovy managed infrastructure the compute is billed through Tensovy on a single invoice, with the hourly cost shown before anything starts. With bring your own key, your provider bills your account directly and Tensovy adds nothing to that rate. What you pay Tensovy covers the orchestration software either way, and is quoted per workload during early access.
Should I use managed infrastructure or my own cloud account?
Managed is the shorter path: there is no cloud account to set up and the compute arrives on one bill with the rest of your usage. Bring your own key suits you if you already have a provider relationship, because existing credits and negotiated rates apply exactly as they do today and we add nothing to your provider's rate. Both routes are in early access, and moving between them changes nothing about how you train or what you get back.
What stops a run from burning money?
You approve a cost and duration estimate before anything launches, and every infrastructure action after that is gated on your approval. An hourly price cap applies, the instance is shut down with one click, and the meter stays in view the whole time.
Do I pay per token for inference?
No. Today you download the weights and run them wherever you like, so inference costs whatever your own hardware or cloud provider charges. A serverless endpoint is on the roadmap.
Training and models
What the agent can build and how it recovers.
What happens on a CUDA out of memory error?
The agent catches most of these before the long run. It smoke tests the package on the GPU first, reads any traceback, patches the config or code and reruns until the test passes. Automatic retuning and checkpoint resume during a long run are on the roadmap.
Which model families are supported?
Open weight language models in the Llama and Qwen families with LoRA and QLoRA through Unsloth, Qwen2.5 VL for document tasks, and detection models in the YOLO family through LibreYOLO, trained on your own labelled images.
Can I override the agent's choices?
Yes. The plan is shown before launch, and you can ask the agent to change the base model or any hyperparameter before you approve it. Every generated file, including the config, stays readable, so nothing about the run is hidden from you.
Can I retrain on fresh data later?
Yes. Every run is recorded in the model registry with its config, dataset and metrics, so retraining on new data is a new session from the same recipe pointed at the updated dataset. Scheduled retraining is on the roadmap.
Data and ownership
Where your data goes and what you keep.
What if I have no formatted dataset?
Tensovy trains on datasets that are already labelled and clean, so it does not convert raw documents or generate labels for you. Upload CSV, JSONL or a labelled image folder, or search Hugging Face and Kaggle from inside the workspace, and the Dataset Studio previews and profiles what you bring before a GPU starts. Dataset preparation from raw documents is on the roadmap.
Can I use public datasets?
Yes. Hugging Face dataset IDs and Kaggle datasets are supported as direct inputs, including mixing them with your own data.
Where does my training data live?
Uploads and imports are staged in Tensovy's object storage under your account so the agent can push them to the training instance. The run itself happens on Tensovy managed infrastructure or on an instance in your own cloud account, and you can delete a dataset at any time.
Can I take the model and leave?
Yes. Adapters and merged models download as safetensors, and detectors export as PyTorch weights. The weights are yours, so nothing stops working when your Tensovy access does.
Is my provider key safe?
A key only comes into it on bring your own key, because managed infrastructure needs no cloud account from you. When you do connect one, the key is stored server side, never sent to the model or to the training instance, and used only to provision, inspect and terminate the resources a run requires. Team access controls are on the roadmap.
Question not covered here?
Ask it on the form. It goes to the people building this, not a support queue, and you get an answer in a day or two.