Amazon Bedrock fine-tuning for classification: what it takes, and how to skip the plumbing

Your team has categorized thousands of tickets, emails, claims or documents. You'd like an AI model that categorizes the next ones the same way, and your company would rather it ran in your own AWS account. Amazon Bedrock can do that: it fine-tunes Amazon Nova models on your examples and serves the result, pay-as-you-go, in your account.

This post explains how Bedrock fine-tuning works, what it costs, when it's worth it, and the parts that take the time: the training file, the IAM roles, the deployment, and testing whether the result is any good. Then it shows how StayCharted does those parts for you, from a spreadsheet, so someone who has never opened the Bedrock console can do it.

A base model like Amazon Nova knows language in general. Fine-tuning trains it further on your own examples, each one a question and the answer you want, until it gives your answers by itself. For classification, every example is a record (a ticket, an email, a description) and the category your team gave it. The result is a custom model: your copy of Nova, tuned to your categories, that you call like any other Bedrock model.

AWS’s ATIS example shows the gap: on an airline-request dataset (ATIS), base Nova Micro got 41.4% right; after fine-tuning, it got 97% on the held-out test set. Your results depend on your data. This is one AWS experiment, not a StayCharted result or a guarantee. Compare prompting and fine-tuning on the same held-out examples to choose for your task.

1. Pick a base model. For text, Bedrock fine-tunes Amazon Nova Micro, Nova Lite, Nova Pro and Nova 2 Lite, in US East (N. Virginia).
2. Write the training file. One JSON object per line (JSONL), in Bedrock's conversation format: a system prompt, the record as the user message, and your category as the assistant's reply. The system prompt you train with must be the one you ask with later.
3. Put it in S3 and give Bedrock a service role that can read it.
4. Start a fine-tuning job and choose hyperparameters: epochs (1 to 5 for Nova, default 2), learning rate, warm-up steps. A job runs for an hour or more.
5. Deploy the custom model for on-demand inference. Nova custom models can answer pay-as-you-go, at the applicable Nova custom-model inference rate, without buying Provisioned Throughput. You call it with the deployment's ARN as the model ID.
6. Test it on examples it never saw, and decide whether it's good enough to use.

All of it is on your AWS bill:

• Training: charged by the number of tokens trained on (records × epochs). AWS's ATIS example, about 1.75 million tokens over 3 epochs, cost $2.18 and took about an hour and a half.
• Storage: a monthly fee for each custom model you keep; $1.75 a month in AWS's example.
• Answers: per token, at the applicable Nova custom-model inference rate.

Short classification records can make a pilot inexpensive, but cost depends on the base model, token count and epochs. Include S3 and any other AWS charges. Check current Amazon Bedrock rates.

The training job is only part of the work. A team also needs to handle:

• Turning a spreadsheet into JSONL, with a system prompt that lists your categories, every row in the same format, nothing malformed (a bad line fails the job).
• Holding back test rows so you can measure the model honestly, and keeping them out of training.
• IAM: a service role for Bedrock, an S3 bucket it can read, permissions for whoever starts the job, all in the right region.
• Waiting and watching: polling the job, then deploying, then waiting for the deployment to be Active.
• Testing: calling the model on every held-out row, comparing answers, working out accuracy per category.
• Deciding what to trust. Bedrock returns an answer, not a confidence, so you need your own rule for which answers a person should check.
• Doing it again every time your team's categories or decisions change.

StayCharted AI Model Trainer does all of the above from a spreadsheet, while the model stays in your AWS account:

1. Connect your AWS account once. In StayCharted, Settings → AI platforms → Amazon Bedrock, tick Also train models here, and launch the CloudFormation template it gives you in us-east-1. The template makes a role for StayCharted, a private bucket for training files and Bedrock's service role. You paste back one role ARN. There are no keys to copy.
2. Upload your examples: a spreadsheet of past records and the category your team gave each one. StayCharted checks the data first: rows that look like they belong in another category, the same text under two categories, near-duplicates.
3. Choose where it trains: Train a model on your examples, then Your servers — Amazon Bedrock, and a Nova base (Micro, Lite or Pro).
4. StayCharted does the rest: holds back test rows, writes the JSONL with the same prompt it will ask with, puts it in your bucket, starts Bedrock's fine-tuning, deletes the file when training ends, deploys the tuned model for on-demand answers, and tests it on the held-out rows.
5. Read the report: accuracy per category and the examples it got wrong, the same report as every other StayCharted model.
6. Use it from a spreadsheet fill, the API, Zapier, Power Automate, Google Sheets, or Claude and ChatGPT through StayCharted's connector. Answers come from your Bedrock; unsure ones go to the Review Queue on eligible plans, and the corrections go into the next version.

You never write JSON, open the IAM console, or call an API.

• Text only. Fine-tuning on your Bedrock is for text models. For pictures, use StayCharted's picture models.
• One region. Bedrock fine-tunes Nova in us-east-1, so the connection used for training has to be there.
• Confidence is estimated. A tuned Nova model gives an answer, not a probability. StayCharted estimates how sure it is from how your team labelled the most similar past examples, and sends anything below a high bar to review.
• Quotas. Bedrock limits how many fine-tuning jobs an account can run at once. A busy account may need a quota increase.
• Each retrain is a new custom model, with its own small monthly storage fee. Delete versions you no longer need in the Bedrock console.

Checked October 22, 2026.

• AWS: supported fine-tuning models and regions
• AWS: Nova fine-tuning example and ATIS results
• AWS: custom model on-demand deployment
• AWS: model hyperparameters
• AWS: training data format
• Amazon Bedrock pricing