Export Training Data

Export the query log as a filtered training dataset for fine-tuning open-source models.

Prerequisites

A Xilos account with admin access
Query history (at least 100 queries recommended)
An understanding of which routing rules or models you want to fine-tune

Why Export Training Data?

Every query that passes through Xilos is logged with full context — the query, the response, the routing rule, the model used, and quality scores. This data is already a training dataset. By filtering for high-quality interactions, you can create a training set that captures your organization's specific query patterns and desired responses.

Step 1: Review Suggestions

Before exporting, check the Suggestions page for fine-tune recommendations:

Navigate to SuperAdmin > Suggestions (if available) or use the API.
Review the suggested fine-tune targets — these are routing rules where fine-tuning would save the most money.
Note the estimated savings and recommended base model.

Step 2: Preview the Training Data

Use the API to preview the training data before exporting:

curl "https://api.xilos.ai/api/v1/training-data/preview?limit=5" \
  -H "Authorization: Bearer YOUR_XILOS_API_KEY"

Review the preview to ensure the data quality meets your standards.
Adjust filters as needed.

Step 3: Export the Dataset

Export the training data with quality filters:

curl -X POST https://api.xilos.ai/api/v1/training-data/export \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_XILOS_API_KEY" \
  -d '{
    "format": "openai",
    "min_faithfulness": 0.7,
    "min_relevance": 0.7,
    "limit": 10000
  }'

Filter Parameters

Parameter	Description	Default
`format`	Output format: `openai`, `alpaca`, `chatml`, `csv`	`openai`
`min_faithfulness`	Minimum faithfulness score (0-1)	0.7
`min_relevance`	Minimum answer relevance score (0-1)	0.7
`routing_rule_id`	Filter by specific routing rule	None
`llm_id`	Filter by specific model	None
`start_date`	Filter from date	None
`end_date`	Filter to date	None
`limit`	Maximum number of examples	10000

Step 4: Review the Output

The export returns a JSON response with:

examples — Array of training examples in the selected format
total_count — Number of examples exported
estimated_tokens — Total token count in the dataset
estimated_cost — Estimated fine-tune cost
format — The output format used

OpenAI JSONL Format

Each line is a JSON object:

{"messages": [{"role": "system", "content": "You are a helpful assistant."}, {"role": "user", "content": "What is the capital of France?"}, {"role": "assistant", "content": "The capital of France is Paris."}]}

Step 5: Create a Fine-Tune Job

Use the exported data to create a fine-tune job:

curl -X POST https://api.xilos.ai/api/v1/fine-tune/create \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_XILOS_API_KEY" \
  -d '{
    "base_model": "meta-llama/Llama-3.3-70B-Instruct",
    "provider": "together",
    "training_data_export_id": "uuid-from-step-3"
  }'

See Fine-Tune a Model for the full workflow.

Info: The estimated fine-tune cost is typically a fraction of what you would spend on premium API calls. A dataset of 10,000 examples might cost $5-20 to fine-tune, but save hundreds per month in API costs.

Best Practices

Filter by quality scores (faithfulness ≥ 0.7, relevance ≥ 0.7)
Filter by specific routing rules to create targeted fine-tune datasets
Export at least 1,000 examples for meaningful fine-tuning
Review the preview before exporting to catch quality issues
Use the OpenAI JSONL format for broadest compatibility
Re-export periodically as your query log grows

Enable Context Compression Fine-Tune a Model