How TensorTurn Works: From Dataset Upload to Deployed API

Here's how TensorTurn works end to end: you import a dataset (CSV, Excel, a zip/rar/7z archive, or a URL), run automated health checks, describe the model you want in plain English so TensorTurn writes and runs the training notebook on cloud or your own GPUs, and then flip one switch to deploy the best weights as a live /predict API. It supports tabular and image tasks across scikit-learn, XGBoost, PyTorch, Keras, and YOLO/Ultralytics.

TensorTurn is designed so a domain expert can drive the whole machine-learning lifecycle without writing training code. Every stage — data import, cleaning, health checks, training, and deployment — happens in one place, and the AI does the heavy lifting while keeping each step reviewable. This page walks through the workflow from upload to a deployed model.

The workflow at a glance

• 1. Import your data — CSV, Excel (.xlsx), zip/rar/7z archives, or from a URL.
• 2. Run automated dataset health checks and fix issues with the AI data editor.
• 3. Describe the model in plain English; TensorTurn generates and runs the training notebook on isolated GPUs.
• 4. (Optional) Pool your own GPUs to train faster or larger via ensemble or fused mode.
• 5. Deploy the best weights as an authenticated /predict API endpoint.

Step 1 — Import your data

Bring data in as a CSV, a multi-sheet Excel file, an archive (zip, rar, or 7z), or straight from a URL. For image datasets, TensorTurn auto-detects common layouts — ImageFolder, YOLO, COCO, and VOC — so you don't have to reformat. Uploads are subject to a zip-bomb safety cap to keep the platform safe.

Step 2 — Check and clean the data

TensorTurn runs automated health checks before you train. For tabular data it looks for train/test leakage, outliers, suspicious correlations, and mixed types, then gives a quality score and a preprocessing playbook. For images it finds exact and near-duplicates (perceptual hash), blur, bad exposure, corrupt files, and resolution problems; it flags cross-split leakage; and it surfaces likely-mislabeled images using DINOv2 embeddings on GPU and a kNN vote, complete with a suggested label and confidence. The AI data editor is a Google-Sheets-style grid: you tell it what to change and it plans reviewable, whitelisted transforms with one-click autofix. For privacy, only a column profile and a few example values leave your browser — never the full dataset — and you can download the cleaned file. You can also share aggregate-only, revocable dataset health reports via a public link.

Step 3 — Chat to build and train

Describe the model you want in plain English — the task, the target, the kind of model. TensorTurn generates the entire training notebook and runs it on isolated cloud GPUs (Modal T4, L4, or A10G) using the right framework for the job: scikit-learn or XGBoost for tabular, PyTorch or Keras for deep learning, YOLO/Ultralytics for detection. If a run errors, self-healing kicks in: the AI patches only the broken cells and retries — up to 100 attempts — learning from each verified fix instead of restarting from scratch.

Step 4 — Scale across your own GPUs (optional)

You can connect your own GPU with a one-line command on Windows, macOS, or Linux. The agent is outbound-HTTPS-only, so it works behind any NAT, firewall, or VPN without opening ports, and your token is stored as a SHA-256 hash. Live telemetry shows CPU, RAM, GPU, VRAM, temperature, and power. Pool several of your machines into one run two ways: ensemble mode, where each machine trains a diverse model on the full data and they vote, or fused mode, where the data is sharded and weights are periodically averaged (DiLoCo-style) and merged into one final model. A work-stealing scheduler keeps machines busy, and a dropped machine's work is requeued.

Step 5 — Deploy as an API

When you're happy with a model, save the best weights and flip one switch to get a live, authenticated /predict endpoint that scales to zero on Modal. It serves YOLO, scikit-learn, PyTorch, ONNX, and Keras models. Authentication is a Bearer API key (stored as a SHA-256 hash), every call is logged, and an inference playground gives you ready-to-paste cURL, JavaScript, Python, and Rust snippets.

Supported data and frameworks

• Tasks: tabular (classification, regression) and image.
• Frameworks: scikit-learn, XGBoost, PyTorch, Keras, YOLO/Ultralytics.
• GPUs: Modal-hosted T4, L4, A10G — or your own connected machines.
• Data formats in: CSV, Excel (.xlsx), zip/rar/7z, URL; image layouts ImageFolder/YOLO/COCO/VOC.
• Deployable model formats out: YOLO, scikit-learn, PyTorch, ONNX, Keras.