Mannequin customization transforms general-purpose AI fashions into specialised enterprise belongings. By fine-tuning basis fashions (FMs) on domain-specific knowledge, companies train AI their distinctive workflows, terminology, and deep area specialization, together with strict adherence to model voice and fewer hallucinations. For enterprises, that is greater than an optimization. It’s the creation of proprietary mental property. A fine-tuned mannequin encodes a company’s distinctive intelligence and finest practices into its structure. This builds a aggressive benefit that’s tough to copy with off-the-shelf public frontier fashions. On the identical time, fine-tuning smaller, open-weight fashions on focused duties usually matches or exceeds the efficiency of a lot bigger proprietary fashions. This strategy delivers important value financial savings whereas conserving delicate knowledge inside safe, personal infrastructure.
Amazon SageMaker AI presents a wide array of open supply fashions and fine-tuning methods to assist organizations tailor basis fashions to their distinctive wants. Now, SageMaker AI introduces serverless model customization for NVIDIA Nemotron 3 fashions, beginning with Nemotron 3 Nano (30B complete parameters, 3B energetic) and Nemotron 3 Tremendous (120B complete parameters, 12B energetic). With supervised fine-tuning (SFT), reinforcement studying with verifiable rewards (RLVR), and reinforcement studying with AI suggestions (RLAIF), you’ll be able to adapt these high-performance open-weight fashions to your particular domains and workflows with out provisioning or managing any infrastructure. For a whole record of open fashions obtainable for serverless mannequin customization, see Customize open weight models in the Amazon SageMaker AI documentation.
On this submit, we discover what makes the Nemotron 3 structure distinctive, stroll by the fine-tuning methods obtainable, and present you step-by-step easy methods to get began with serverless customization utilizing SageMaker Studio.
Overview of NVIDIA Nemotron 3 fashions on Amazon SageMaker AI
NVIDIA Nemotron 3 is a household of open-weight massive language fashions (LLMs) constructed on a hybrid Mamba-Transformer Combination-of-Specialists (MoE) structure with native help for as much as 1M-token context lengths. The structure interleaves three complementary layer sorts: Mamba-2 layers for environment friendly linear-time sequence processing, Transformer consideration layers for exact associative recall, and Latent Combination-of-Specialists (LatentMoE) layers that compress tokens earlier than routing to specialised specialists. This design prompts solely a fraction of complete parameters per ahead move (for instance, 12B of 120B within the Tremendous variant), delivering excessive throughput and powerful accuracy at considerably decrease compute value. The fashions use multi-environment reinforcement studying by NeMo Fitness center, which aligns them to real-world, multi-step agentic duties throughout domains akin to coding, reasoning, and long-context evaluation.
Nemotron 3 Nano 30B
Nemotron 3 Nano is a small language mannequin optimized for prime compute effectivity whereas sustaining robust accuracy on specialised duties. Nemotron 3 Nano performs strongly on coding and reasoning duties amongst open language fashions in its dimension class. Educated utilizing multi-environment reinforcement studying by NeMo Gym, the mannequin achieves 4x larger throughput than its predecessor Nemotron 2 Nano. Its environment friendly 3B energetic parameter footprint makes it supreme for high-volume, multi-agent workloads the place value and latency matter. For a deeper have a look at the structure and coaching methods, see the NVIDIA developer blog.
Nemotron 3 Tremendous 120B
Nemotron 3 Tremendous is a bigger mannequin designed for high-efficiency multi-agent AI and complicated reasoning duties that require extra capability than Nano whereas sustaining value effectivity. Nemotron 3 Tremendous delivers excessive compute effectivity, throughput, and accuracy for complicated multi-agent purposes akin to software program growth and cybersecurity triaging. The mannequin performs effectively at reasoning, coding, and long-context evaluation, whereas remaining environment friendly sufficient to run repeatedly at scale. This makes it a great match for IT ticket automation, enterprise workflow orchestration, and autonomous agent programs that require sustained multi-step reasoning. For extra particulars, see the NVIDIA developer blog on Nemotron 3 Super.
SageMaker AI serverless mannequin customization
Amazon SageMaker AI serverless mannequin customization removes the undifferentiated heavy lifting of fine-tuning. You don’t have to provision GPU clusters, configure distributed coaching frameworks, or handle checkpointing and fault tolerance. SageMaker AI handles infrastructure provisioning and coaching orchestration, so you’ll be able to focus in your knowledge, enterprise use case, and analysis, and pay just for what you utilize. You’ll be able to be taught extra about SageMaker AI serverless model customization in the AWS documentation.
For Nemotron 3 fashions, SageMaker AI serverless mannequin customization helps the Supervised High-quality-Tuning (SFT), Reinforcement Studying with Verifiable Rewards (RLVR) and Reinforcement Studying from AI Suggestions (RLAIF) fine-tuning methods.
| Approach | Description | Finest For |
| Supervised High-quality-Tuning (SFT) | Present labeled input-output pairs to show the mannequin new behaviors. | Excessive-quality examples of the habits you need: area Q&A pairs, formatted device calls, style-aligned responses, or task-specific instruction completions |
| Reinforcement High-quality-Tuning (RFT / RLVR) | Use Reinforcement Studying with Verifiable Rewards (RLVR) to optimize mannequin habits in opposition to a reward sign. The mannequin generates a number of candidate responses per immediate, a reward perform scores them, and the mannequin updates its coverage to favor what works. | Duties with naturally verifiable targets like device calling accuracy, code correctness, or format compliance |
| Reinforcement Studying from AI Suggestions (RLAIF) | Use a separate AI mannequin to information the mannequin optimization. An AI mannequin evaluates mannequin outputs and gives suggestions indicators, which helps iterative coverage enchancment with out human-labeled reward knowledge. | Aligning mannequin tone, helpfulness, and security; enhancing response high quality when human analysis is dear or subjective; refining open-ended era duties |
Let’s stroll by easy methods to get began with serverless mannequin customization for Nemotron 3 fashions. Whereas the bottom Nemotron 3 fashions ship robust general-purpose efficiency, enterprise use instances want domain-specific habits that base fashions alone can not obtain. With mannequin customization, you’ll be able to adapt these fashions for industry-specific terminology and choice patterns, practice dependable device calling along with your group’s APIs, align outputs along with your model voice, refine multi-step agentic reasoning in your architectures, and optimize value by specializing the smaller Nano mannequin to match bigger mannequin efficiency on focused duties.
Getting began with SageMaker AI serverless mannequin customization
You will get began with serverless mannequin customization by the Amazon SageMaker Studio console or programmatically utilizing the SageMaker Python SDK. On the console, navigate to the Fashions web page, choose your Nemotron 3 mannequin, and comply with the guided workflow to configure your coaching knowledge and launch a customization job. Alternatively, should you’re already working inside SageMaker AI, you need to use the agentic functionality with agent skills to speed up your mannequin customization workflow. The next sections stroll you thru the conditions, knowledge preparation, and step-by-step directions utilizing the SageMaker Studio console. For an in depth programmatic instance with the SageMaker Python SDK for customizing an open-source mannequin, see the AWS samples GitHub repository.
Stipulations
Earlier than you start, confirm that you’ve:
- An AWS account with AWS Identification and Entry Administration (IAM) permissions for Amazon SageMaker AI.
- A SageMaker AI area with Studio entry.
- Your coaching knowledge within the required construction and format.
Put together your coaching knowledge for SageMaker AI serverless mannequin customization
Excessive-quality coaching knowledge is the muse of any profitable fine-tuning job. For serverless mannequin customization on SageMaker AI, your knowledge have to be formatted as JSONL (JSON Strains), the place every line represents a single coaching instance. The particular schema will depend on the method you select: SFT requires conversation-format examples with labeled input-output pairs, whereas RFT (RLVR) requires prompts paired with floor fact values in your reward perform. Correctly structured knowledge ensures the mannequin learns the behaviors you propose with out introducing noise or formatting errors. For a hands-on walkthrough of getting ready your coaching knowledge, see the Data Preparation module in the SageMaker AI serverless model customization workshop. Alternatively, if you’re working with SageMaker AI, you need to use the built-in coding agent with agent skills to automatically prepare and validate your knowledge formatting, decreasing guide effort and serving to you get to coaching sooner.
Mannequin customization in SageMaker AI Studio
Observe these steps to customise a Nemotron 3 mannequin utilizing the SageMaker AI Studio console.
- Open Amazon SageMaker AI Studio and within the left navigation pane, select Fashions.
- Navigate to the mannequin you wish to customise within the UI. Seek for “NVIDIA” to search out the Nemotron 3 household of fashions, and choose the NVIDIA mannequin that you really want (
NVIDIA-Nemotron-3-Nano-30B-*orNVIDIA-Nemotron-3-Tremendous-120B-*) for the subsequent step.
- Choose your mannequin customization method from the supported Supervised High-quality-Tuning (SFT), Reinforcement Studying with Verifiable Rewards (RLVR) and Reinforcement Studying from AI Suggestions (RLAIF) fine-tuning methods.
When selecting a reward perform sort for RLVR, contemplate your job necessities. The built-in reward perform (Actual Match, Code Execution, Math Solutions) works effectively for duties with single, objectively right solutions, requiring no further code. Select a customized reward perform when your job wants richer scoring logic, akin to partial credit score, format checks, reasoning high quality analysis, or domain-specific guidelines. With customized reward capabilities, you’ll be able to rating on a number of indicators, form rewards to keep away from all-zero gradients on early rollouts, emit observability metrics, and encode the Python verification logic your job requires. For detailed steerage on authoring and registering a customized reward perform, see the RLVR workshop documentation. - Configure your coaching knowledge by deciding on an current dataset (if obtainable) or creating a brand new dataset (see the previous part for details about getting ready your dataset).
- Set the customization hyperparameters or use advisable defaults.
- Select Submit to launch the mannequin customization job.
SageMaker AI routinely provisions the required compute, executes the coaching job, and captures steady logs. The coaching metrics are routinely logged to the SageMaker MLflow App by default for coaching monitoring.
Monitor coaching progress
You’ll be able to monitor the standing on the mannequin house web page, which shows coaching efficiency, as proven within the following screenshot. A number of high-level metrics are value monitoring. Practice Reward (for RLVR) ought to improve steadily. Coaching Loss and Validation Loss ought to lower and monitor generalization, respectively. Coverage Entropy (for RLVR) decreases because the mannequin beneficial properties confidence. Gradient Norm ought to stabilize to point convergence.
The detailed coaching and validation metrics are additionally logged to the related SageMaker AI MLflow App, as proven within the following screenshot. This captures a complete set of metrics and parameters that monitor coaching progress, and mannequin habits. Within the MLflow monitoring UI, these metrics are organized by the element they measure (actor, critic, rollout, efficiency), so you’ll be able to diagnose coaching well being at a look.
Consider your fine-tuned mannequin
After coaching completes, you’ll be able to consider the fine-tuned mannequin utilizing the built-in analysis options of SageMaker AI serverless mannequin customization. It gives three strategies to evaluate the standard of your custom-made mannequin, as proven within the following screenshot. LLM-as-a-Decide makes use of an Amazon Bedrock frontier mannequin to grade responses in opposition to high quality metrics with out requiring ground-truth labels. Customized Scorer applies your individual reward capabilities or built-in scorers to provide normal pure language processing (NLP) metrics akin to F1, ROUGE, and BLEU. Benchmarks scores your mannequin on standardized educational benchmarks (MMLU, BBH, GPQA, MATH, IFEval) for broad functionality evaluation throughout reasoning, data, and instruction-following.
You can too activate Examine with base mannequin in analysis to instantly measure how your post-trained mannequin performs relative to the bottom mannequin. Along with the earlier coaching metrics, MLflow tracks the coaching dynamics (rewards, KL divergence, loss). The analysis measures output high quality from an end-user perspective, supplying you with a whole image of the mannequin fine-tuning effectiveness.
Deploy the fine-tuned mannequin
Deploy your custom-made mannequin instantly from the mannequin particulars web page on the console. You can too deploy to SageMaker Inference endpoints, or you’ll be able to obtain mannequin weights from an Amazon Easy Storage Service (Amazon S3) bucket for self-managed deployment. The deployment choices auto-populate defaults, supplying you with full flexibility over compute and scaling based mostly in your visitors and throughput necessities. The next screenshot exhibits the deployment of the fine-tuned NVIDIA Nemotron Nano 30B utilizing an ml.g6e occasion powered by NVIDIA L40S Tensor Core GPUs. The deployment makes use of SageMaker inference parts and, by default, serves the merged mannequin weights, the place the base model and LoRA adapter are mixed right into a single set of weights for optimized inference. As a result of it is a LoRA fine-tune, you may as well self-host and serve the unmerged LoRA adapter individually, as a result of you might have entry to each the bottom weights and the adapter weights in your S3 bucket. After deployment, you invoke the endpoint utilizing the invoke technique with the AWS Command Line Interface (AWS CLI) or SDK.
Clear up
To keep away from incurring pointless prices, we advocate deleting your SageMaker AI Studio domain, SageMaker Endpoints, and some other sources that you simply created after you’re performed utilizing them. The particular value of utilizing SageMaker AI serverless mannequin customization will depend on the bottom mannequin you select and the customization stage. See the Amazon SageMaker AI pricing page for the price breakdown and particulars.
Conclusion
With serverless mannequin customization for NVIDIA Nemotron 3 fashions on Amazon SageMaker AI, now you can adapt these high-performance open-weight fashions to your particular domains and workflows. Whether or not you’re fine-tuning Nemotron 3 Nano for cost-efficient agentic job execution or customizing Nemotron 3 Tremendous for complicated multi-agent orchestration, SageMaker AI handles compute provisioning, coaching orchestration, and metric monitoring so you’ll be able to focus in your knowledge, analysis, and deployment.
Get began right now with serverless Model Customization on Amazon SageMaker AI. For detailed examples of customizing open-source fashions, see the AWS samples GitHub repository. To be taught extra, see the Amazon SageMaker AI model customization documentation.
Concerning the authors
