Gemini Fine-Tune JSONL Validator

updated 17 May 2026 · Vertex AI supervised tuning · separate pages for OpenAI, Anthropic, Llama, Mistral

Google Gemini fine-tune JSONL validator. Validates Vertex AI supervised-tuning data: contents array, parts[], user/model alternation, optional systemInstruction, function-calling tools, multimodal fileData and inlineData, and the 131,072-token-per-example cap. Catches Vertex's exact consecutive_turns and Invalid role 'assistant' errors before the dataset validation job rejects your file.

Your training data never leaves this tab. Vertex AI pulls your dataset from GCS when a job starts; this pre-flight check is fully local.

⌨ Prefer the terminal? jsonlkit validate --gemini data.jsonl — same checks, in a pipe.

Where can I fine-tune Gemini in 2026?

One platform: Vertex AI supervised tuning on Google Cloud. The free tuning that used to be available in Google AI Studio / Generative Language API was deprecated in May 2025 with Gemini 1.5 Flash-001, and no replacement was launched — if you find a tutorial showing the {"text_input": "...", "output": "..."} AI Studio shape, that's the old format and the endpoint that consumed it no longer exists.

Supported Gemini models for supervised tuning (Vertex AI, May 2026):

• Gemini 2.5 Pro — text, image, audio, video, document, function-calling.
• Gemini 2.5 Flash — same modality set.
• Gemini 2.5 Flash-Lite — text-focused; cheaper.
• Gemini 2.0 Flash / 2.0 Flash-Lite — scheduled for shutdown on 1 June 2026. Don't start new SFT jobs.

Older 1.x models, including 1.5 Pro and 1.5 Flash, are no longer tunable on either Vertex AI or the deprecated AI Studio endpoint.

The exact JSONL shape Vertex AI accepts

Text — canonical example

{
  "systemInstruction": {
    "role": "system",
    "parts": [{"text": "You are a precise legal summarizer."}]
  },
  "contents": [
    {"role": "user",  "parts": [{"text": "Summarize Clause 4.2."}]},
    {"role": "model", "parts": [{"text": "Clause 4.2 limits liability to fees paid in the prior 12 months."}]}
  ]
}

Function calling — full round-trip

Function-calling tuning is a supported modality on the 2.5 family. The training example shows the model how to emit a functionCall, how to consume a functionResponse, and how to compose the final answer.

{
  "contents": [
    {"role": "user",  "parts": [{"text": "Weather in Kyiv tomorrow?"}]},
    {"role": "model", "parts": [{"functionCall": {"name": "get_weather", "args": {"city": "Kyiv", "when": "tomorrow"}}}]},
    {"role": "user",  "parts": [{"functionResponse": {"name": "get_weather", "response": {"tempC": 14, "sky": "rain"}}}]},
    {"role": "model", "parts": [{"text": "14 C with rain."}]}
  ],
  "tools": [{"functionDeclarations": [{
    "name": "get_weather",
    "description": "Get weather by city and time",
    "parameters": {
      "type": "object",
      "properties": {"city": {"type": "string"}, "when": {"type": "string"}},
      "required": ["city"]
    }
  }]}]
}

Multimodal — image via GCS URI

For dataset-scale multimodal tuning, Vertex AI prefers fileData pointing at a GCS URI over inline base64. Use inlineData for small one-offs.

{"contents": [
  {"role": "user",  "parts": [
    {"fileData": {"mimeType": "image/jpeg", "fileUri": "gs://my-bucket/cat.jpg"}},
    {"text": "What breed?"}
  ]},
  {"role": "model", "parts": [{"text": "Maine Coon."}]}
]}

The deprecated AI Studio shape — don't use this

{"text_input": "...", "output": "..."}
// Old Google AI Studio / Generative Language API tuning format.
// Endpoint deprecated May 2025. Will NOT work on Vertex AI.

Vertex AI dataset limits

What this validator checks

• Each line is valid JSON, one object per line.
• Top-level contents is present and is a non-empty array.
• Optional systemInstruction is an object with its own parts array.
• First contents entry role is user.
• Roles strictly alternate user / model; no assistant anywhere (a common copy-paste mistake from OpenAI samples).
• Every parts is a non-empty array.
• Every part has exactly one of text, functionCall, functionResponse, inlineData, or fileData.
• Optional tools is an array of functionDeclarations objects with proper JSON-Schema parameters.
• Per-example token estimate flagged if it exceeds 131,072.

Common mistakes this validator catches

{"contents":[{"role":"assistant","parts":[{"text":"broken"}]}]}
// Error: Gemini uses 'model', not 'assistant'.

{"contents":[{"role":"user","parts":{"text":"Hi"}}]}
// Error: 'parts' must be an array, not an object.

{"contents":[{"role":"user","content":"Hi"}]}
// Error: Gemini uses 'parts', not 'content'.

{"contents":[
  {"role":"user","parts":[{"text":"A"}]},
  {"role":"user","parts":[{"text":"B"}]}
]}
// Error: consecutive_turns — roles must alternate user/model.

{"text_input":"Q","output":"A"}
// Error: deprecated AI Studio shape; not accepted by Vertex AI.

Real Vertex AI error strings

How to use it

1. Drop a .jsonl file or paste records directly.
2. Click Validate. Every line is parsed and every part is checked.
3. Inspect the Error List — each entry maps to a Vertex error you'd see at dataset-validation time.
4. Download valid examples only to ship a clean file to GCS.
5. Upload to GCS, then start the tuning job from the Vertex AI console or gcloud ai CLI.

Tips & common pitfalls

• The role is model, not assistant. This is the single most-copy-pasted mistake on the web — tutorials lifted from OpenAI examples and not relabelled.
• Content lives in parts, always as an array. Even a single text message: "parts": [{"text": "..."}].
• System instruction goes at the top level. Inside contents, only user and model are allowed.
• Use fileData for dataset-scale multimodal. Inline base64 (inlineData) inflates file size and pushes you past token limits quickly; GCS URIs are streamed.
• Function-calling tuning needs the top-level tools array. Same JSON-Schema shape as runtime function calling — Vertex re-uses it for SFT.
• Don't mix Vertex and AI Studio shapes in the same file. The dataset validator will reject the whole file.

Troubleshooting

Frequently asked questions

Related tools

• OpenAI Fine-Tune ValidatorChat messages with tool_calls, vision, developer role.
• Anthropic Fine-Tune ValidatorBedrock Claude 3 Haiku — top-level system + alternating messages.
• Llama Fine-Tune ValidatorLlama 3 chat, ShareGPT, Alpaca.
• Mistral Fine-Tune ValidatorOnline alternative to validate_data.py.
• Fine-tune dataset editorVisual editor; export to any provider format.
• Token counterPer-example count against the 131,072 cap.