Anthropic (Claude) Fine-Tune JSONL Validator

updated 11 May 2026 · Bedrock supervised tuning format

Your training data never leaves this tab. Bedrock uploads your file when a job starts; this pre-flight check is fully local — useful when the dataset contains anything you would rather not upload twice.

Anthropic's shape, in one line

{"system": "You are helpful.", "messages": [
  {"role": "user",      "content": "Hi"},
  {"role": "assistant", "content": "Hello"}
]}

The two non-obvious rules that trip everyone up:

• System is a top-level string, not a message. Putting {"role":"system",...} inside messages is the #1 error this validator catches.
• Turns must strictly alternate user → assistant → user → assistant… starting with user. Two user turns in a row, or starting with assistant, is rejected.

What the validator checks

• Each line is valid JSON.
• Top-level messages is present and is an array.
• Optional system is a string if present.
• No message has role: "system" (system goes at the top level).
• First message role is user; last is assistant.
• Roles strictly alternate user / assistant.
• Each content is either a string or an array of typed blocks (text, image).
• Image blocks have a valid source with type (base64 / url) and media_type.

How to use it

1. Drop your .jsonl file into the dashed box, or paste the text directly.
2. Click Validate. Every line is parsed and every message is checked.
3. If errors are found, look at the Error List for the line number and reason.
4. Click Download valid examples only to save just the lines that passed.

Common mistakes I catch

{"messages": [
  {"role": "system", "content": "..."},
  {"role": "user",   "content": "Hi"}
]}
// Error: Anthropic puts system at the top level, not in messages.

{"messages": [
  {"role": "user", "content": "Hi"},
  {"role": "user", "content": "Two user turns"}
]}
// Error: Roles must alternate user/assistant.

{"messages": [
  {"role": "assistant", "content": "I start the conversation."}
]}
// Error: First message must be user.

Multi-modal content

Vision messages use a typed array of content blocks:

{"system": "You describe images.", "messages": [
  {"role": "user", "content": [
    {"type": "image", "source": {"type": "base64", "media_type": "image/jpeg", "data": "..."}},
    {"type": "text",  "text": "What is this?"}
  ]},
  {"role": "assistant", "content": "A cat."}
]}

Frequently asked questions

Related tools

• OpenAI Fine-Tune ValidatorChat (messages) and legacy prompt/completion formats.
• Gemini Fine-Tune ValidatorVertex AI contents+parts, user/model roles.
• Llama Fine-Tune ValidatorLlama 3 chat and the ShareGPT conversations shape.
• Mistral Fine-Tune ValidatorMistral API messages array and instruct templates.
• Fine-tune dataset editorVisual editor for chat-style rows with one-click export.
• Train / Val / Test splitterSplit a clean file 80/10/10 before training.