Dataset Templates for Fine-Tuning
Stop wondering "what should my training data look like?" These 6 JSONL templates have example rows in the exact format you need. Replace with your data and start training.
6 Templates, 6 Use Cases
Each template includes 10 example rows showing the exact format, a consistent system prompt, and structured outputs you can adapt to your domain.
Product Classifier
Visual product identification for retail, inventory, and POS systems. Trains a model to classify items from image descriptions into structured categories.
Format: JSONL (chat completions)
Examples: 10 example rows
Use case: Retail checkout, inventory scanning, warehouse sorting
Document Extractor
Structured extraction from messy documents — invoices, receipts, forms, shipping labels. The model learns to output clean JSON from unstructured text.
Format: JSONL (chat completions)
Examples: 10 example rows
Use case: Invoice processing, form digitization, receipt scanning
Domain Chatbot
Specialized Q&A for customer support, knowledge bases, or internal tools. Multi-turn conversation format with consistent tone and helpful responses.
Format: JSONL (chat completions)
Examples: 10 example rows
Use case: Support bots, internal knowledge assistants, FAQ automation
Text Classifier
Categorization of text into custom taxonomies — feedback routing, ticket triage, content moderation. Outputs structured JSON with category and confidence.
Format: JSONL (chat completions)
Examples: 10 example rows
Use case: Ticket routing, feedback analysis, content categorization
Code Assistant
Framework or codebase-specific coding assistant. Trains a model on your project's conventions, patterns, and API surface so it gives contextually accurate guidance.
Format: JSONL (chat completions)
Examples: 10 example rows
Use case: Internal copilot, codebase onboarding, documentation Q&A
Tone Adapter
Brand voice and style matching. Given generic text, rewrites it in your specific brand voice. Learns sentence length, vocabulary, personality, and formatting preferences.
Format: JSONL (chat completions)
Examples: 10 example rows
Use case: Marketing copy, email drafts, social media, documentation
From Template to Trained Model
Pick a template
Choose the template closest to your use case. Each one uses the chat completions JSONL format that Ertas (and most fine-tuning platforms) expect.
Replace with your data
Swap the example rows with your own data. Keep the same structure — system prompt, user input, assistant output. Aim for 100-500 examples minimum.
Upload and train
Upload your JSONL file to Ertas, choose a base model, and start training. No code, no YAML configs, no CLI. The visual interface handles everything.
Download and deploy
Download your fine-tuned model as a GGUF file. Run it locally with llama.cpp, Ollama, or any GGUF-compatible runtime. $0 per inference, forever.
Dataset Quality Checklist
Before you upload, make sure your dataset hits these marks.
Each row follows the same system prompt / user / assistant structure
Minimum 100 examples (500+ recommended for complex tasks)
No duplicate rows — each example teaches something different
Assistant outputs are consistent in format (always JSON, always markdown, etc.)
Edge cases are represented — not just the happy path
User inputs vary in length, phrasing, and complexity
No personally identifiable information (PII) unless intentional
File is valid JSONL — one JSON object per line, no trailing commas
Ready to Fine-Tune?
Get all 6 templates and start building your dataset today. Upload to Ertas when you're ready — no ML expertise needed.