Amazon Nova recipes - Amazon Nova

Amazon Nova recipes

You can get an Amazon Nova recipe from the recipes repository. An Amazon Nova recipe is a YAML configuration file that provides details to SageMaker on how to run your model customization job. It provides the base model name, sets training hyperparameters, defines optimization settings, and includes any additional options required to fine-tune or train the model successfully.

You can also access Amazon Nova recipes through SageMaker AI Monarch and by navigating to the model hub, selecting AWS, and browsing Amazon Nova models to find their associated recipes. Both SageMaker AI Monarch and provide sample notebooks for each recipe, which include all the necessary steps to modify the recipes and run customization jobs using SageMaker training jobs or SageMaker HyperPod environments.

To access the recipes pages in Amazon SageMaker Studio, the execution role must have the following permissions.

JSON
{ "Version":"2012-10-17", "Statement": [ { "Effect": "Allow", "Action": [ "s3:GetObject" ], "Resource": [ "arn:aws:s3:::*model-customization-recipes*" ] } ] }

To execute the sample notebooks on SageMaker training jobs and , use one of the following SageMaker distribution image versions: 2.7.1+, 2.8.0+, 3.2.1+, 3.3.0+. This applies to both SageMaker AI Monarch and .

Getting Amazon Nova recipes

To get a base Amazon Nova recipe, clone the recipes repository by running the following command.

git clone https://github.com/aws/sagemaker-hyperpod-recipes.git

The base recipes are available at recipes_collection/recipes/.

cd recipes_collection/recipes/

The Amazon Nova customization recipes are in the following folders.

Recipe type Folder
SFT (Full-rank and PEFT) fine-tuning/nova
Evaluation evaluation/nova
CPT training/nova

Getting Amazon Nova Forge recipes

To access specialized Amazon Nova Forge recipes for jobs, please set up your access to Amazon Nova Forge and then follow this workshop to set up the Forge version of the CLI. For SageMaker Training Jobs, the AWS Console and Nova Customization SDK will automatically access Forge recipes.

Available models and algorithms

The following table summarizes the availability of customization for Amazon Nova models and supported algorithms using SageMaker.

Model name

Model ID

Fine-tuning

Notes

Amazon Nova Lite 2.0

amazon.nova-2-lite-v1:0:256k

Yes

  • SFT - Accepts either text and/or image or text and/or video as the input and produces text as output. A single job can't combine text, image, and video in the same run.

  • RFT - Accepts text and image as input for single-turn conversations and produces text as output. Improved reward optimization capabilities.

  • CPT

Amazon Nova 2.0 Lite

The table below lists detailed information of the Amazon Nova 2.0 Lite recipes reference.

Model Category/Sub-category Technique Recipe Name Image URI (Training Jobs) Image URI (SageMaker HyperPod) Compute Instance
Nova 2.0 Lite Training Continued Pre Training nova_lite_2_0_p5x8_gpu_pretrain.yaml N/A 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-HP-CPT-V2-latest ml.p5.48xlarge, ml.p5en.48xlarge
Nova 2.0 Lite Parameter Efficient Fine Tuning Fine Tuning nova_lite_2_0_p5_gpu_lora_sft.yaml 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-TJ-SFT-V2-latest 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-HP-SFT-V2-latest ml.p5.48xlarge, ml.p5en.48xlarge
Nova 2.0 Lite Full-Rank Fine Tuning Fine Tuning nova_lite_2_0_p5_gpu_sft.yaml 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-TJ-SFT-V2-latest 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-HP-SFT-V2-latest ml.p5.48xlarge, ml.p5en.48xlarge
Nova 2.0 Lite Parameter Efficient RFT Fine Tuning nova_lite_v2_smtj_p5_p5en_gpu_lora_rft.yaml 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-TJ-RFT-V2-latest 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-HP-RFT-TRAIN-V2-latest ml.p5.48xlarge, ml.p5en.48xlarge
Nova 2.0 Lite Full Rank RFT Fine Tuning nova_lite_v2_smtj_p5en_gpu_rft.yaml 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-TJ-RFT-V2-latest 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-HP-RFT-TRAIN-V2-latest ml.p5.48xlarge, ml.p5en.48xlarge
Nova 2.0 Lite Bring Your Own Data Evaluation nova_lite_2_0_p5_48xl_gpu_bring_your_own_dataset_eval.yaml 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-TJ-Eval-V2-latest 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-HP-Eval-V2-latest ml.p5.48xlarge, ml.p5en.48xlarge
Nova 2.0 Lite General Text Benchmark Evaluation nova_lite_2_0_p5_48xl_gpu_general_text_benchmark_eval.yaml 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-TJ-Eval-V2-latest 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-HP-Eval-V2-latest ml.p5.48xlarge, ml.p5en.48xlarge
Nova 2.0 Lite RFT Evaluation Evaluation nova_lite_2_0_p5_48xl_gpu_rft_eval.yaml 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-TJ-Eval-V2-latest 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-HP-Eval-V2-latest ml.p5.48xlarge, ml.p5en.48xlarge
Nova 2.0 Lite LLM Rubric Based Judge Evaluation nova_lite_2_0_p5_48xl_gpu_rubric_llm_judge_eval.yaml 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-TJ-Eval-V2-latest 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-HP-Eval-V2-latest ml.p5.48xlarge, ml.p5en.48xlarge