Amazon Nova Micro Fine-Tuned Text-to-SQL Models Now Available on Bedrock On-Demand Inference at $0.80/Month for 22,000 Q

Amazon Nova Micro Fine-Tuned Models Available on Bedrock On-Demand Inference for Text-to-SQL

Amazon Web Services has announced that fine-tuned Amazon Nova Micro models can now be deployed on Bedrock's on-demand inference infrastructure for custom text-to-SQL generation. According to AWS testing, a sample workload of 22,000 queries per month incurred costs of $0.80 monthly, compared to higher costs with persistent model hosting infrastructure.

The solution applies LoRA (Low-Rank Adaptation) fine-tuning to Nova Micro, enabling organizations to customize the model for proprietary SQL dialects and domain-specific database schemas while maintaining serverless, pay-per-token pricing.

Technical Implementation

AWS provides two implementation paths for fine-tuning Nova Micro:

Bedrock Model Customization: Fully managed fine-tuning through the AWS console or API, with training data uploaded to S3. AWS handles underlying infrastructure and the resulting custom model deploys with the same token-based pricing as base Nova Micro with no additional markup.

SageMaker AI Training Jobs: Provides granular control over hyperparameters and training infrastructure for organizations requiring customization beyond managed options.

Both approaches use the same data preparation pipeline and deploy to Bedrock for on-demand inference.

Training Configuration

The demonstration uses the sql-create-context dataset, combining WikiSQL and Spider datasets with over 78,000 examples of natural language questions paired with SQL queries. Training data is formatted as JSONL files with system prompts, user queries, and SQL responses.

Configurable hyperparameters for Nova Micro fine-tuning:

• Epochs: 1-5 (AWS used 5 in testing)
• Batch Size: Fixed at 1 for Nova Micro
• Learning Rate: 0.000001-0.0001 (AWS used 0.00001)
• Learning Rate Warmup Steps: 0-100 (AWS used 10)

Training completion time: approximately 2-3 hours according to AWS.

Infrastructure Requirements

Deployment requires:

• AWS account with billing enabled
• IAM permissions for Bedrock Nova Micro, SageMaker AI, and Bedrock Model Customization
• Quota for ml.g5.48xl instance for SageMaker AI training

Amazon Bedrock automatically generates training and validation loss metrics, stored in S3. AWS reports that successful training shows both losses decreasing consistently and converging to comparable final values.

What This Means

The on-demand inference option removes the primary cost barrier to deploying fine-tuned models for specialized use cases. Organizations with variable text-to-SQL workloads can now customize models for proprietary SQL dialects without maintaining persistent infrastructure. The $0.80/month cost figure for 22,000 queries demonstrates viability for production workloads with intermittent usage patterns, though AWS does not disclose baseline costs for comparison or specify whether this includes only inference costs or total end-to-end expenses. The LoRA approach trades higher per-query latency for zero idle costs, making it suitable for applications where sub-second response times are acceptable.