Tool
CSV/TSV to instruction JSONL
Map header columns to instruction, input, and output, then write one JSON object per row. The table never leaves this tab.
Parsing runs in this tab. A file you pick is read by the browser; GearDel does not receive it.
How a row becomes JSONL
The first row is the header. Later rows are records. Quoted fields follow RFC 4180: a doubled quote is a literal quote, and commas or tabs inside quotes stay in the cell.
Delimiter can be comma, tab, or semicolon. Auto picks whichever of those appears most often on the first logical line, outside quotes.
You map columns to instruction, optional input, and output. Known header names (instruction, prompt, 지시문, output, 응답, …) are guessed; you can override them.
Alpaca keeps instruction, input, and output. ShareGPT and OpenAI messages join instruction and input with a blank line into one user turn, then one assistant turn. That is one pair per spreadsheet row, not multi-turn chat.
A file chosen from disk is decoded as UTF-8 unless you pick EUC-KR. EUC-KR uses the same WHATWG index as the byte-count page (web euc-kr, windows-949). Bytes with no pointer are dropped and counted, not guessed.
What this page does not do
- It does not implement Excel formulas, merged cells, or .xlsx. Export CSV or TSV first.
- It does not invent a Korean morphological split of one cell into instruction and output.
- Tool-call and image messages are not written. Use the JSONL checker if you already have those lines.
- At most 20,000 data rows or about two million characters.