Bedrock Model Fine-tuner

Bedrock Model Fine-tuner is a helper library for fine-tuning Amazon Bedrock models. This toolkit assists in generating Q&A datasets from documents and streamlines the LLM fine-tuning process.

Features

• Wrapper classes for easy fine-tuning and deployment of Bedrock models using boto3 (currently limited to Claude 3 Haiku model)
• Q&A dataset generation from documents for Claude 3 Haiku model fine-tuning
• Dataset validation to ensure compliance with Claude 3 Haiku model fine-tuning format and constraints

Usage

Step 1: Generate Q&A Dataset from Documents (Optional)

import boto3
from core import ChatModelId, QaDatasetGenerator, get_llm

boto_session = boto3.Session(region_name="us-west-2")

llm = get_llm(
    ChatModelId.CLAUDE_V3_5_SONNET,
    region_name="us-west-2",
)
qa_dataset_generator = QaDatasetGenerator.from_jsonl(llm, "../assets/docs.jsonl")

train_dataset = qa_dataset_generator.generate(dataset_type="train")
_ = qa_dataset_generator.save_and_upload(
    train_dataset,
    "../assets/train_dataset.jsonl",
    boto_session=boto_session,
    bucket_name="<YOUR-S3-BUCKET-NAME>",
)

Step 2: Validate Q&A Dataset

from core import QaDatasetValidator

qa_dataset_validator = QaDatasetValidator()
qa_dataset_validator.validate_data("../assets/train_dataset.jsonl")

Step 3: Fine-tune and Deploy Model

from core import BedrockModelFinetuner

bedrock_model_finetuner = BedrockModelFinetuner(aws_region_name="us-west-2")

_ = bedrock_model_finetuner.finetune("s3://<YOUR-S3-BUCKET-NAME>/datasets/train_dataset.jsonl")
_ = bedrock_model_finetuner.deploy()

Step 4: Delete Provisioned Model (Optional)

_ = bedrock_model_finetuner.delete()

For detailed usage examples, please refer to the notebook files in the samples directory.

Additional Resources

• Amazon Bedrock Custom Models Documentation
• AWS Samples
• Fine-tune Anthropic's Claude 3 Haiku in Amazon Bedrock to Boost Model Accuracy and Quality

License

MIT License