Fine-tune, evaluate, and deploy AI models on Amazon SageMaker.
npx skills add https://github.com/aws/agent-toolkit-for-aws --skill aws-ai-ml
Please help me install this Agent Skill. Skill: aws-ai-ml Source: https://github.com/aws/agent-toolkit-for-aws/tree/main/plugins/aws-core/skills/aws-ai-ml Command: npx skills add https://github.com/aws/agent-toolkit-for-aws --skill aws-ai-ml
Customizing and deploying foundation models on AWS involves many decisions — choosing a base model, picking a fine-tuning technique, validating datasets, generating training code, and managing endpoints. This Skill guides AI coding agents through the entire SageMaker model customization lifecycle with validated workflows, so nothing critical gets missed.
Ask the agent to help you fine-tune a model on SageMaker, for example: "Help me fine-tune a Qwen model on my dataset and deploy it to a SageMaker endpoint."
High-intent search queries and answers about installing and using this skill.
Select a base model from SageMaker Hub, choose a fine-tuning technique (SFT, DPO, RLVR, or RLAIF), validate your JSONL training dataset, and generate training code from the provided templates. The workflow handles IAM roles, S3 buckets, and EULA acceptance before starting the training job.
This Skill has validated workflows for SFT (Supervised Fine-Tuning), DPO (Direct Preference Optimization), RLVR (Reinforcement Learning from Verifiable Rewards), and RLAIF. Technique availability depends on the selected model and is validated against the model's recipes.
Yes, both OSS models (Llama, Mistral, Qwen) and Nova models fine-tuned through SageMaker Serverless Model Customization can be deployed to SageMaker endpoints or Bedrock. JumpStart base models deploy only to SageMaker real-time endpoints.
All training and evaluation datasets must be serialized as JSONL (one JSON object per line) and match the schema for your model type and fine-tuning technique. Non-JSONL files like parquet or CSV must be transformed first using the dataset transformation workflow.
The endpoint diagnostics workflow collects endpoint status via DescribeEndpoint, CloudWatch metrics (invocations, errors, latency, utilization), and recent container logs. It is read-only and covers inference issues, not training job failures.
It does not cover Ground Truth labeling, Feature Store, HyperPod training, full fine-tuning (FFT), BYO container training, traditional ML models like XGBoost, or general-purpose AWS infrastructure. For those cases it offers best-effort guidance from general AWS knowledge.
awsDiagnose AWS Clean Rooms permission and logging issues across IAM, S3, KMS, and Lake Formation.
frankxaiCoordinate Bedrock, SageMaker, and AWS AI services to automate enterprise AI workflows.
Alex1980AlexAutomate Claude Code administration with OpenTelemetry, IAM, and cost analytics.
open-mercatoGuide users through creating and structuring AI Skills with SKILL.md best practices.
