skills.rest
awslabs
awslabsOFFICIAL
Explore all skills by this author →
WebsiteGitHub

finetuning-technique

Selects and validates SFT, DPO, RLVR, or RLAIF fine-tuning techniques against SageMaker model recipes.

881|152|Updated Feb 5, 2026
Search Tags:#model training#fine-tuning#DPO#sft#rlaif#Sagemaker#RLVR
One-click install
npx skills add https://github.com/awslabs/agent-plugins --skill finetuning-technique
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: finetuning-technique
Source: https://github.com/awslabs/agent-plugins/tree/main/plugins/sagemaker-ai/skills/finetuning-technique
Command: npx skills add https://github.com/awslabs/agent-plugins --skill finetuning-technique

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires boto3, and includes scripts (resource) and references (resource) components.

What problem does it solve?

Choosing the wrong fine-tuning technique wastes training time and money, and not every model on SageMaker supports every technique. This Skill recommends the right technique for your use case and verifies it is actually available for your selected base model before you commit.

Core Features & Use Cases

  • Technique Recommendation: Consults a selection guide to recommend SFT, DPO, RLVR, or RLAIF based on whether your task has verifiable answers, subjective quality judgments, or preference data.
  • Recipe Validation: Runs a boto3 script against the SageMaker Hub to confirm the chosen technique is supported by the selected model's available recipes.
  • Use Case: You have selected a Llama base model and want to align it to your brand voice. The Skill recommends RLAIF, checks the model's recipes, and confirms support before you proceed to training configuration.

Quick Start

Ask the assistant to help you choose and validate a fine-tuning technique for your selected base model on SageMaker.

Frequently Asked Questions about finetuning-technique

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I choose between SFT, DPO, RLVR, and RLAIF for fine-tuning?▼

Choose SFT for tasks with clear right answers, DPO when you have preference pairs for subjective quality, RLVR when outputs can be verified programmatically like code or math, and RLAIF when quality is subjective but you lack human preference data.

How do I check which fine-tuning techniques a SageMaker model supports?▼

Run the get_recipes.py script with the model name and hub name. It calls the SageMaker describe_hub_content API and returns the supported techniques filtered to SFT, DPO, RLVR, and RLAIF.

What is the difference between DPO and RLAIF?▼

DPO requires a static dataset of human preference pairs collected upfront. RLAIF uses an AI judge model to generate preference signals dynamically, enabling iterative improvement without human annotators.

When should I not use RLAIF for model fine-tuning?▼

Avoid RLAIF when the task has objectively verifiable answers (use RLVR), when you already have high-quality human preference data (use DPO), or when the AI judge model is weaker than the model being trained.

What happens if my selected model does not support the chosen technique?▼

The Skill explains that the model does not support the technique on SageMaker and offers to return to model selection so you can pick a different base model whose recipes include the technique.

Related Skills

Explore More Software Engineering Skills→
awsaws
2.5k

aws-ai-ml

Fine-tune, evaluate, and deploy AI models on Amazon SageMaker.

Official
Advanced
fanguyunfanguyun

fine-tuning-expert

Automates LLM fine-tuning with LoRA/QLoRA, dataset prep, hyperparameters, and deployment optimization.

Community
Advanced
Strategic-AutomationStrategic-Automation
6

evaluating-llms-harness

Run standardized LLM evaluations across 60+ benchmarks using lm-evaluation-harness.

Official
Intermediate
doanchienthangdevdoanchienthangdev
4

LLMOps

Manages LLM lifecycle including prompts, RAG pipelines, versioning, A/B testing, evaluation, deployment, monitoring, and cost tracking.

Community
Advanced