Back to catalog / Agent workflows

Fine-tune Haiku on Bedrock and host a task model

Validate conversational JSONL, upload it to S3, launch a Bedrock customization job, and host the resulting model for inference.

Goal

Produce a callable customized Claude Haiku model from a validated sample set rather than only inspecting training data.

Prepare JSONL with system/messages: first message user, last assistant, at least two messages, and alternating roles. [src10-anthropic-finetune-haiku-task-model-on-bedrock]
Upload training data to your S3 bucket and submit a customization job using a Bedrock service role allowed to read it. [src10-anthropic-finetune-haiku-task-model-on-bedrock]
Poll job status with boto3; after completion create the model deployment/inference profile and test using held-out prompts. [src10-anthropic-finetune-haiku-task-model-on-bedrock]

Job status is Completed, deployment succeeds, and held-out prompts follow the target JSON format without copying training examples.

For role denial, check S3 read permissions and region; for job failure, check JSONL first/last roles and extra fields before starting a new job.

AWS account, Bedrock customization eligibility/quota, S3 bucket, and a service role with read access; model access may require enablement.

Uploads JSONL to S3 and creates paid training/hosting resources; sample data is sent to AWS Bedrock for processing.

Training time, price, and regions depend on current Bedrock support; use approved, de-identified examples only.

These values describe the upstream repository, not popularity or likes for an individual skill or playbook.

No metrics snapshot is available yet.

License: MIT