Guides

Model Fine-Tuning

Fine-tuning adapts weights on specific task distributions, style constraints, or proprietary enterprise taxonomies. Fine-tuned checkpoints run with dedicated inference guarantees on Cortiqa hardware.

When to Fine-Tune

• Prompt Engineering vs Fine-Tuning: Always test prompt engineering and few-shot examples first. Fine-tune only when prompt instructions cannot reliably enforce style, output syntax, or complex domain rules.
• Cost & Latency Reduction: Ingesting long instructions into weights allows you to shorten input prompts significantly, saving tokens on every subsequent call.
• Steerability: Guarantee strict adherence to custom programming languages, internal DSLs, or clinical report styles.

Training Data Format

Training sets must be formatted as conversational JSON Lines (.jsonl):

{"messages": [{"role": "system", "content": "You are a customer support specialist for FinTech Corp."}, {"role": "user", "content": "How do I dispute a credit card charge?"}, {"role": "assistant", "content": "To dispute a charge, navigate to Statements > Select Transaction > File Dispute within 60 days."}]}
{"messages": [{"role": "system", "content": "You are a customer support specialist for FinTech Corp."}, {"role": "user", "content": "What is the fee for wire transfers?"}, {"role": "assistant", "content": "Domestic outgoing wire transfers incur a flat $15 fee. Incoming wires are free of charge."}]}

Training Lifecycle

1. Dataset Curation: Prepare at least 200 - 1,000 high-quality, verified conversation examples.
2. Upload Dataset: Upload your .jsonl file to the console or via /api/v1/files.
3. Create Job: Select your base model, hyper-parameters (epochs, learning rate multiplier), and start training.
4. Deployment: Once validation loss converges, your custom model suffix (e.g. ft:openai/gpt-oss-120b:org:v1) is ready for immediate deployment.

Best Practices

• Quality over Quantity: 500 meticulously cleaned, accurate examples will outperform 10,000 noisy, duplicate samples.
• Validation Split: Always hold out 10-20% of examples as a test set to monitor against overfitting.
• Enterprise Inquiries: For large-scale proprietary foundational model training, contact team@cortiqa.co.