Fine-Tuning Foundation Models in Bedrock: Customizing AI for Your Needs

gcptutorials.com AWS

Amazon Bedrock's fine-tuning capability lets you adapt foundation models like Claude or Titan to your specific domain without managing infrastructure. This tutorial walks through customizing a text generation model using your proprietary data - perfect for startups and enterprises alike.

Step 1: Prepare Bedrock for Fine-Tuning

Objective: Configure IAM roles and training data storage

Create Training IAM Role

{
    "Version": "2012-10-17",
    "Statement": [
        {
            "Effect": "Allow",
            "Action": [
                "bedrock:CreateModelCustomizationJob",
                "s3:GetObject",
                "s3:PutObject"
            ],
            "Resource": "*"
        }
    ]
}

Validate Bedrock Access

aws bedrock list-foundation-models \
--query "modelSummaries[?byOutputModality=='TEXT']" \
--region us-east-1

Step 2: Format Training Data

Dataset: Custom FAQ pairs in JSONL format

import json

training_data = [
    {"input": "What's return policy?", "output": "30-day returns for unused items"},
    {"input": "Do you price match?", "output": "We match authorized retailers"}
]

with open("train.jsonl", "w") as f:
    for item in training_data:
        f.write(json.dumps(item) + "\n")

Upload to S3

aws s3 cp train.jsonl s3://your-bucket/fine-tuning/data/

Step 3: Configure Fine-Tuning Job

Key Parameters:

training_config = {
    "baseModelIdentifier": "anthropic.claude-v2",
    "customModelName": "support-bot-v1",
    "trainingDataConfig": {"s3Uri": "s3://your-bucket/fine-tuning/data/train.jsonl"},
    "hyperParameters": {
        "epochCount": "5",
        "batchSize": "32",
        "learningRate": "0.0001"
    },
    "outputDataConfig": {"s3Uri": "s3://your-bucket/fine-tuning/output/"}
}

Start Training Job

aws bedrock create-model-customization-job \
--job-name "support-bot-training" \
--role-arn "arn:aws:iam::123456789012:role/BedrockFineTuningRole" \
--cli-input-json file://training_config.json \
--region us-east-1

Step 4: Track Training Progress

CLI Monitoring

aws bedrock get-model-customization-job \
--job-identifier "support-bot-training" \
--region us-east-1

CloudWatch Metrics

aws cloudwatch get-metric-data \
--metric-data-queries '[
    {
        "Id": "loss",
        "MetricStat": {
            "Metric": {
                "Namespace": "AWS/Bedrock",
                "MetricName": "TrainingLoss",
                "Dimensions": [{"Name": "JobName", "Value": "support-bot-training"}]
            },
            "Period": 300,
            "Stat": "Average"
        }
    }
]' \
--start-time $(date -u +"%Y-%m-%dT%H:%M:%SZ" -d "-1 hour") \
--end-time $(date -u +"%Y-%m-%dT%H:%M:%SZ")

Step 5: Deploy Fine-Tuned Model

Create Endpoint

aws bedrock create-provisioned-model-throughput \
--model-id "arn:aws:bedrock:us-east-1::custom-model/support-bot-v1" \
--provisioned-model-name "support-bot-prod" \
--model-units 1 \
--region us-east-1

Python Inference

import boto3

bedrock = boto3.client('bedrock-runtime', region_name='us-east-1')

response = bedrock.invoke_model(
    modelId="support-bot-v1",
    body=json.dumps({
        "prompt": "\n\nHuman: What's your warranty period?\n\nAssistant:",
        "max_tokens_to_sample": 300
    })
)

print(json.loads(response['body'].read())['completion'])

Step 6: Evaluate Model Quality

Automatic Evaluation (CLI)

aws bedrock create-model-evaluation-job \
--job-name "support-bot-eval" \
--model-identifier "support-bot-v1" \
--evaluation-data-config '{"s3Uri": "s3://your-bucket/eval/data.jsonl"}' \
--evaluation-metrics "Accuracy,Precision" \
--region us-east-1

Manual Testing

test_cases = [
    "How do I reset password?",
    "Where's my order confirmation?"
]

for query in test_cases:
    response = bedrock.invoke_model(
        modelId="support-bot-v1",
        body=json.dumps({"prompt": f"\n\nHuman: {query}\n\nAssistant:"})
    )
    print(f"Q: {query}\nA: {json.loads(response['body'].read())['completion']}\n")