What is supervised fine-tuning (SFT)?

SFT trains on curated input-output pairs to reshape a model's behavior, teaching it how to respond—following instructions, adhering to schemas, or adopting a tone—without injecting new knowledge.

Why is data formatting important?

Formatting shapes learned behavior because SFT teaches the model to respond to a specific input shape. AWS recommends a conversational JSONL format with one JSON object per line, strict role alternation, and including system prompts when used during inference.

What quality checks does AWS recommend?

AWS recommends auditing accuracy, ensuring diversity (semantic coverage, information depth), maintaining consistency within similar tasks, deduplicating, and screening for toxicity and bias, with examples cited from LIMA and AlpaGasus.

Large Language ModelsAmazon AI Blog5d agoAug 27, 2026, 04:00 JST2 min read

AWS shows how to prep SFT data for fine-tuning

Send on LINE

Key takeaway

AWS explains how to prepare data for supervised fine-tuning.
It covers quality, formatting, and splits.
High-quality data beats large volumes, and careful curation can match models trained on much more data.

3 Key Points

What happened
AWS published the first part of a two-part series on preparing data for supervised fine-tuning (SFT), covering quality checks, formatting requirements, and train/evaluation splits.
Why it matters
Getting data preparation right determines the ceiling of any SFT project, and the post highlights that quality beats quantity—LIMA showed 1,000 carefully curated examples can match models trained on far more data, and AlpaGasus showed filtering to the cleanest 20 percent can train faster and score higher.
What to watch
The post recommends a conversational JSONL format with strict role alternation, including system prompts and reasoning traces (e.g., using reasoningContent) for reasoning-capable models, with the second post to cover advanced strategies like data mixing and augmentation.

Ask the AI about this article →

Summaries like this, in your inbox every morning.

Get news like this every morning — free

Context & Analysis

This post is part of a series aimed at practitioners who have decided to fine-tune a foundation model, laying the groundwork before advanced strategies like data subset selection and augmentation. It positions data preparation as the key determinant of fine-tuning success, with a clear emphasis on quality over quantity, citing LIMA and AlpaGasus to support this. The guidance is model-agnostic, though it references Amazon Bedrock and Amazon Nova 2.0 for formatting examples, which may encourage adoption of AWS's tooling. The first post's focus on fundamentals—quality checks and formatting—sets the stage for the second post, which will cover more advanced techniques, suggesting a comprehensive approach to SFT data preparation within the AWS ecosystem.

FAQ

What is supervised fine-tuning (SFT)?: SFT trains on curated input-output pairs to reshape a model's behavior, teaching it how to respond—following instructions, adhering to schemas, or adopting a tone—without injecting new knowledge.
Why is data formatting important?: Formatting shapes learned behavior because SFT teaches the model to respond to a specific input shape. AWS recommends a conversational JSONL format with one JSON object per line, strict role alternation, and including system prompts when used during inference.
What quality checks does AWS recommend?: AWS recommends auditing accuracy, ensuring diversity (semantic coverage, information depth), maintaining consistency within similar tasks, deduplicating, and screening for toxicity and bias, with examples cited from LIMA and AlpaGasus.

Amazon AI BlogRead Original Article

Send on LINE

Get the latest Large Language Models news every morning

For example, today's edition would include:

DataAgent launches with $10M to auto-fix Kubernetes faultsSiliconANGLE AI · 46m ago
SK Hynix custom HBM boosts inference up to 5.15xDIGITIMES Asia · 46m ago
Nvidia Earnings: Boring by Design, Avoiding a Consolidated WorldStratechery (Ben Thompson) · 46m ago

AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.

Subscribe free →

Free · takes 30 seconds · unsubscribe anytimeWhat is AIToday? →