Dataset builder: documents to ready-to-train JSONL
Drop your documents and get a ready-to-train JSONL: it reads them, OCRs scans and old-font pages, cuts chunks, lets AI write question-answer pairs and checks every answer against the real document.
Autopilot: documents → ready-to-train JSONL
Drop your documents. It reads them, runs OCR on scans and old-font pages, cuts them into chunks, lets AI write question-answer pairs (within the credit limit you choose), and checks every answer against its passage. If nothing needs changing, just save the .jsonl.
2. Check and clean the records
0 records: 0 question-answer pairs, 0 text chunks · about 0 tokens
Two lists: Q&A pairs are what an AI learns from. Chunks are the document text the pairs were written from — kept so you can check them (and for RAG).
Clean the records
3. Choose the format and save
Token numbers are estimates (no tokenizer runs here); the platform counts exactly. Check a sample of the data before you pay for training.
Versions
A version is a saved copy of the dataset. The autopilot saves one by itself. Save another after you edit, review or update — then see exactly what changed, or open an older one. Kept on this device only.
Nothing to save yet. Drop documents above.
Give it your own name (optional)
No versions yet. The autopilot saves version 1 by itself.
More ways to add material (tables, pasted Q&A, existing .jsonl, AI Q&A by hand)
1. Add your material
PDF, Word, text or Markdown are read here and cut into chunks (keeps file name and page). Good for RAG and embeddings — or let AI write Q&A from the chunks.
Documents and tables are read on your device and are not uploaded. Only if AI writes the Q&A, the text chunks are sent through our server to Google Gemini (not stored). Versions are saved in this browser only.
About this tool
Review side by side, keep versions, compare them, and save for OpenAI, Gemini or RAG. Files are read on your device.