How to use the JSONL Dataset Validator
1. Paste your JSONL content (one JSON object per line) into the input panel — or upload a .jsonl file.
2. Each line is validated against OpenAI's fine-tuning format: messages array with system/user/assistant roles.
3. Invalid lines show the specific error. Valid lines show message count and estimated token count.
Tip: minimum viable fine-tuning dataset is 10 examples; recommended is 50–100+ high-quality examples.
What is it?
JSONL (JSON Lines) is the format used for OpenAI fine-tuning datasets. Each line is a valid JSON object with a messages array — the same structure as the Chat Completions API. Fine-tuning teaches a model to behave in a specific way based on your examples.
This validator checks each line for: valid JSON structure, presence of the messages array, valid roles (system/user/assistant), non-empty content, and the required minimum of one user + one assistant message. It also estimates the token count per example and the approximate training cost at OpenAI's fine-tuning rate ($8/1M tokens).
The file never leaves your browser — all validation is done client-side.
Example
{"messages": [{"role": "user", "content": "Hello"}, {"role": "assistant", "content": "Hi!"}]}✓ Line 1: valid · 2 messages · ~12 tokens
Each line must be a self-contained conversation. The model learns from all examples collectively.
Benefits
Generate accurate code from your actual data
TypeScript interfaces, Python dataclasses, and OpenAI function schemas are generated directly from your real JSON — accurate field names, correct types, no guesswork.
Control LLM API costs before you scale
Token counts and per-call cost estimates for GPT-4o, Claude 3.5, and Gemini let you optimise prompt size and model choice before a feature goes to production.
Ready-to-run code and configs, not boilerplate
Every tool produces output you can paste directly into your project — API request code, prompt templates, chunk configurations — adapt rather than write from scratch.
Purpose-built for AI engineering workflows
From RAG chunking strategy to prompt template selection to function calling schema design — tools built for the specific problems AI engineers encounter daily.
Frequently asked questions
How many examples do I need for fine-tuning?
OpenAI requires a minimum of 10 examples but recommends 50–100+ for meaningful improvement. Quality matters more than quantity — 50 excellent examples outperform 500 mediocre ones.
How is fine-tuning cost calculated?
OpenAI charges approximately $8 per 1M training tokens (as of early 2026). One training epoch processes every example once, so cost = total_tokens × epochs × price_per_token. The default is 1–3 epochs depending on dataset size.
Can I have multiple turns in one example?
Yes — each example can have multiple user/assistant turns, plus an optional system message at the start. Multi-turn examples help the model learn conversation patterns and context handling.
What format does OpenAI fine-tuning require?
Each line must be: {"messages": [{"role": "system"|"user"|"assistant", "content": "..."},...]}. The messages array must contain at least one user message and one assistant message. System messages are optional.
Should I include a system message in every example?
Not necessarily. Include a system message if you want the model to always use a specific persona or instruction. If your examples have varied system prompts, the fine-tuned model learns to follow whatever system prompt it's given at inference time.