Datasets, structured.
Roboflow, CVAT, Label Studio, FiftyOne, pandas, DVC, a dozen bash scripts — training AI today means fighting a different tool at every stage. DataLith unifies importing, exploring, cleaning, labeling, versioning, augmenting and exporting into one application.
Genesis + Phase 2 · runs entirely in your browser · nothing uploaded anywhere
The problem is fragmentation
A typical dataset workflow means a different tool at every stage — and none of them talk to each other.
14 modules, one long build
This is the whole vision, numbered as a real pipeline — import through export. Some are live today, some are honest previews. The in-studio Roadmap page tracks this same list as it grows.
Universal Importer
Drag & drop, file picker, and full client-side .zip extraction. Cloud connectors (Drive, S3, Hugging Face…) are roadmap.
Visual Dataset Explorer
Fast searchable/filterable grid with an inspector panel. Embedding-space navigation is roadmap.
AI Dataset Understanding
Rule-based answers ship for free and instantly. Connect a BYOK provider in AI Settings and open-ended questions go to a real model, grounded in your real dataset stats.
Automatic Cleaning
Real SHA-256 duplicate detection, corrupt-image decoding checks, empty-file detection, filename normalization.
Dataset Version Control
Linear commit history with diffs and rollback, in-session. Branching, merge, and on-disk storage are roadmap.
AI-Assisted Labeling
Manual classification labeling ships free. Connect a vision-capable BYOK provider and "AI-suggest unlabeled" runs the real label-20/AI-labels-the-rest loop, capped per batch, always shown as "needs review" until you confirm.
Dataset Analytics
Live stat cards, format breakdown, and size-distribution charts computed off your real dataset.
Visual Data Augmentation
Real Canvas 2D pipeline: rotate, flip, brightness, contrast, blur, noise. Weather/perspective effects are roadmap.
Synthetic Data Generation
Generative augmentation of the dataset itself. Needs a model backend — not attempted in this milestone.
Training Integration
One-click PyTorch/HF/YOLO training. Needs a real compute backend — out of scope for a browser tab.
Model Evaluation
Automatic weak-class and bias detection after training. Depends on Module 10 existing first.
Dataset Marketplace
Community publishing and discovery. Needs real backend infrastructure and moderation.
AI Dataset Agent
A rule-based assistant ships free. With a BYOK key connected, unmatched questions go to your chosen model (OpenAI, Anthropic, Google, Groq, OpenRouter, Mistral, DeepSeek, or xAI).
Visual Workflow Builder
Chain the real steps above into an ordered pipeline and run it for real. Grows as more modules land.
Dataset Sourcing
Real, working fetches from Hugging Face's dataset hub and Wikipedia's REST API (both CORS-open, work on the web build), plus paste-in text/HTML — every candidate passes a free heuristic screen and, with a BYOK provider connected, an AI review/curation pass before you can commit it to the dataset. Fetching arbitrary blog/wiki URLs works from the Tauri desktop build (native fetch_url, no CORS); the web build tries direct CORS first and otherwise tells you honestly to use the desktop app or paste instead.
An operating system, not another labeling tool
Support every data type — images, video, audio, text, PDFs, CSV, JSON, medical, 3D, point clouds, time series, synthetic — inside one coherent surface, instead of stitching a dozen single-purpose tools together for every project.
Frontend
- React 19 / Next.js (App Router)
- TypeScript
- Tailwind CSS v4
Desktop
- Tauri (Rust shell)
- Static export target
State
- Zustand
- Client-first, in-memory
Data / ML
- JSZip (import/export)
- Chart.js (analytics)
- BYOK AI (8 providers)