Fine-tuning JSONL formatter and validator

Paste data or open a file, then press Validate & format.

• openai-chat: each line one object with a non-empty "messages" array; roles limited to system, developer, user, assistant, tool; content a string or an array of text parts; optional numeric weight.
• openai-completions: each line one object with string (or string-array) "prompt" and "completion"; optional "suffix" and weight.
• Duplicate keys keep JSON.parse semantics — last value wins — and the loss ledger says so. Number spellings normalize (1.0 → 1) and that is reported too, with counts and the first line it happened on.
• Unknown top-level keys are preserved verbatim, named in the ledger — the formatter never drops your data, and never invents fields a target expects.

• Chat fine-tuning — clean

  Two valid records — one with a training weight.

• Chat fine-tuning — with errors

  Line 2 has an unknown role; line 3 is not JSON — both are named, not dropped silently.

• Legacy completions

  Prompt/completion pairs — the legacy fine-tuning shape.

• JSON array in, JSONL out

  A pretty JSON array becomes one record per line — the wrapper removal is reported.