Training Scenario Overview

Scenario 1: One-Click Training

The simplest entry point designed by ModelArts for beginners, quick prototype validation, and teaching demonstrations. You do not need to write code or prepare complex configurations; simply select classical algorithms pre-built on the platform or popular LLMs (such as the Qwen3 series or DeepSeek series), pair them with datasets provided by the platform or uploaded by you, and start training with a single click. For details, see Creating a One-Click Training Job.

Target audience: Ideal for zero-foundation users, students, and product managers to quickly validate ideas, or for scenarios where users want to see training results within minutes. For users who wish to quickly experience the benefits of Ascend compute, they can directly select the Ascend supernode resource pool to achieve significant performance improvements with almost no parameter modifications.

Scenario 2: Custom Training

The most flexible and widely applicable traditional training method on ModelArts. Create a custom model job when one-click training cannot meet your requirements, for example, when you need to use in-house developed model architectures, custom loss functions, special data augmentation logic, or introduce specific third-party dependency libraries. In this scenario, you need to write the complete training code yourself and organize it according to the directory structure and input/output specifications agreed upon by the platform. You can choose to use preset AI engines (such as PyTorch, TensorFlow, or MindSpore) or custom images to fulfill special environment dependencies. This scenario gives developers maximum flexibility and control, making it the most common approach for production-grade model training. For details, see Creating a Custom Training Job.

Custom training supports both single-node and distributed training.

Target audience: Ideal for algorithm engineers or R&D teams with independent algorithm R&D requirements who need full control over the training process. It is particularly suitable for custom model development scenarios in computer vision, natural language processing, traditional machine learning, etc. It also supports large-scale distributed training on Ascend supernodes, making it the preferred route for most intermediate and advanced developers.

Scenario 3: Model Fine-Tuning

When you do not need to train a model from scratch, but instead want to fine-tune an open-source or built-in LLM (such as DeepSeek or Qwen) using your industry or business data to better adapt it to specific domains or tasks, you should choose fine-tuning jobs. Fine-tuning jobs usually encapsulate the training workflows and best-practice configurations of mainstream LLMs (such as LoRA, full-parameter fine-tuning, and other fine-tuning methods). You only need to prepare a fine-tuning dataset that meets the format requirements, select the base model and fine-tuning strategy, and quickly obtain a customized model that "understands your business." This scenario represents the most frequent training demand in the current LLM era, achieving significant domain adaptation effects at a lower compute cost. For details, see Creating a Fine-Tuning Training Job.

Target audience: Suitable for enterprises or developers who need domain adaptation, supervised fine-tuning (SFT), or instruction-following optimization on foundation models in vertical domains (such as healthcare, finance, law, and customer service). When training LLMs with 10 billion parameters or more, you are advised to use fine-tuning jobs combined with Ascend supernodes to significantly reduce communication overhead, improve training efficiency, and lower overall costs.

Creating Training Jobs via API/SDK

When manual selection on the console is insufficient and you prefer to submit training programmatically, or need to embed training into CI/CD pipelines, scheduled tasks, and automated scheduling systems, you should choose to create training jobs via API, SDK, or command-line tools. This scenario defines all parameters of a training job (code paths, hyperparameters, resource specifications, node counts, etc.) via code, making the training process reproducible, versionable, and batch-schedulable. It is highly suitable for teams requiring frequent experimentation, multi-task concurrency, or building end-to-end MLOps pipelines. It serves as the key entry point for truly "engineering" and "automating" ModelArts training capabilities. For details, see Creating a Training Job Using an API.

Creating training tasks via API/SDK supports both single-node and distributed training.

Target audience: Suitable for MLOps engineers, enterprise R&D teams, or production environments that need to implement automated "data → training → evaluation → deployment" pipelines. It is particularly applicable to scenarios requiring programmatic control, such as periodic training, hyperparameter search, A/B testing, and large-scale experiment management. It can be deeply integrated with the training-inference shared pool scheduling capabilities of Ascend supernodes to achieve efficient resource utilization with automated night-time training and daytime inference.

Quick Reference for Choosing Among Four Scenarios