FINE-TUNING
BETATrain models that speak your SQL.
◇
The Workflow
From raw schema to fine-tuned model in four steps.
workflow
CURATE
Generate from schema
EVOLVE
Expand breadth & depth
EVAL
Test accuracy
EXPORT
JSONL or CSV
Limerence handles steps 1-3. You bring step 4 to your training provider of choice.
◇
Export Formats
Ready for OpenAI, Anthropic, or any training pipeline.
formats
JSONL (OpenAI format)
{"messages": [
{"role": "system",
"content": "You are..."},
{"role": "user",
"content": "Top 10..."},
{"role": "assistant",
"content": "SELECT..."}
]}
CSV (Universal)
question,sql,context
"Top customers?",
"SELECT name...",
"customers table..."
"Monthly sales?",
"SELECT SUM...",
"orders table..."
- •System prompts included
- •Schema context embedded
- •Validation status flagged
◇
Train With Your Provider
Take your exported data to any fine-tuning platform.
compatible with
OpenAI
Fine-tune API
Anthropic
Claude Fine-tune
Anyscale
Llama Training
Together AI
Modal
Your Infra
Standard formats work everywhere. No lock-in.
◇
Why Fine-tune?
When prompting hits its limits.
comparison
PROMPTING
✓Quick to iterate
✓No training required
✓Works with any model
✗Context window limits
✗Higher per-query costs
✗Less consistent
FINE-TUNING
✓Higher accuracy
✓Faster inference
✓Lower token costs
✓Custom behavior baked in
✗Requires training data
✗Model-specific
✗Takes time to train
Start with prompting. Graduate to fine-tuning when you need:
- •95%+ accuracy requirements
- •High query volumes (cost matters)
- •Domain-specific patterns that prompts miss
◇
Iteration Loop
Fine-tuning isn't one-and-done. Use evals to improve.
continuous improvement
┌──────────────────────────────────────┐
│ │
▼ │
EXPORT→TRAIN→EVAL─────┘
│ │
│ ┌───────────────────┘
│ ▼
└──────→IMPROVE
- •Eval reveals blind spots
- •Add failing cases to dataset
- •Re-export and retrain
- •Repeat until accuracy targets met