tensorturnFree Dataset Health Check: Find Data Problems Before You Train
TensorTurn's dataset health check is a free tool that scans your CSV, Excel workbook, or zipped image dataset and returns a full report on the problems that quietly wreck models: train/test leakage, duplicates, outliers, mixed types, corrupt or blurry images, and likely-mislabeled examples. It also gives you a 0-100 quality score and a concrete preprocessing playbook so you know exactly what to fix first. You run it in the browser, no code and no credit card, on the free tier.
What the dataset health check tool does
Most model failures trace back to the data, not the algorithm. The dataset health check automates the checklist a careful ML engineer runs by hand before training. Point it at a tabular file or an image dataset and it profiles every column or image, flags the issues that inflate metrics or degrade accuracy, and ranks them by severity. Instead of discovering leakage or duplicate rows after you've burned GPU hours on a model that looks great in validation and falls apart in production, you see the problems up front.
What it checks: tabular vs image datasets
| Dataset type | What the health check finds |
|---|---|
| Tabular (CSV / Excel) | Train/test leakage, duplicate rows, outliers, highly correlated columns, mixed data types, missing values, plus a 0-100 quality score and a preprocessing playbook |
| Image (zip / folder / URL) | Exact and near-duplicates via perceptual hashing, blurry images, over- and under-exposed images, corrupt files, odd resolutions, cross-split leakage, and likely-mislabeled images with a suggested label and confidence |
Image datasets are auto-detected whether they arrive as ImageFolder, YOLO, COCO, or VOC. The likely-mislabeled and duplicate detection use DINOv2 embeddings computed on a GPU, so semantic near-duplicates and label errors are caught even when two files are not byte-identical.
How it works in four steps
- Import your data: drag in a CSV, an Excel workbook (multi-sheet supported), a zip/rar/7z of images, or paste a URL.
- The tool auto-detects the format and structure and runs every relevant check in parallel.
- You get a ranked report: each issue with severity, the affected columns or images, and why it matters.
- Act on it: follow the preprocessing playbook, use the AI data editor to autofix, or go straight to chatting a model into existence.
The quality score and preprocessing playbook
For tabular data the report ends with a single 0-100 quality score that summarizes how training-ready the dataset is, plus a preprocessing playbook: an ordered list of the specific steps that would raise that score, such as removing leaked columns, capping outliers, or fixing mixed-type fields. It turns a vague worry that the data might be bad into a to-do list you can actually work through.
Why run a health check before training
A duplicate row that lands in both train and test, or a column that is a near-perfect proxy for the label, produces validation accuracy that looks fantastic and collapses on real inputs. Blurry, corrupt, or mislabeled images cap how good any image model can get. Catching these in seconds is far cheaper than training, deploying, and then debugging a model that never had a chance. The health check makes the invisible problems visible before they cost you time and GPU credits.
Share the report, keep the raw data private
You can publish any report as a shareable public link so teammates or reviewers can see the findings. Shared reports are aggregate-only: they contain the statistics and issue summaries, never the raw rows or full-resolution images, and you can revoke the link at any time.
Ready to try it? Upload a dataset and get your free health report at https://www.tensorturn.com/login. Free tier, no credit card required.
Frequently asked questions
Is the dataset health check really free?
Yes. It runs on the free tier with no credit card, subject to the free-tier GPU-hour and upload limits. TensorTurn is currently in beta, and Pro (Rs 899/mo) raises the limits if you need more headroom.
What file formats can I upload?
Tabular: CSV and multi-sheet Excel (.xlsx). Images: zip, rar, or 7z archives, or a URL. Image datasets in ImageFolder, YOLO, COCO, and VOC layouts are detected automatically.
Does my data stay private?
Analysis runs on isolated cloud infrastructure, and image checks use GPU embeddings, so image files are processed in the cloud. Any report you choose to share is aggregate-only, contains no raw rows, and is revocable. The separate AI data editor keeps tabular data in your browser, sending only a column profile and a few example values.
How long does a health check take?
It depends on dataset size. Small tabular files return in seconds; large image datasets that need GPU embeddings for every file take a few minutes.
What should I do after I get the report?
Work through the preprocessing playbook, use the AI data editor to apply reviewable autofixes, and then describe the model you want in plain English to train it on cloud GPUs, all in the same platform.