UI Overview¶

Welcome to the DeepExtension User Interface (UI) — a web-based, modular workspace designed for both domain experts and AI developers to collaborate seamlessly throughout the model development lifecycle.

This section provides a practical overview of all key UI components and how they work together to enable end-to-end AI workflows without requiring users to write code.

What You'll Find in This Section¶

The User Guide is organized by feature areas of the DeepExtension interface. It helps you understand what each part of the UI does and how to use it effectively:

Dashboard¶

Get a real-time overview of active projects, training jobs, model deployments, usage stats, and recent activity — all in one glance. The dashboard includes:

Core Data Statistics Overview: Centralized display of key metrics across models, datasets, knowledge bases, and evaluations
Fine-tuning Task Panel: Bar chart visualization of model training task distribution by type
Base Model Distribution: Donut chart showing proportion of base models in Customized Modelss
Model Assessment Analysis: Area chart displaying evaluation mode distribution across assessment tasks
Third-party Model Ecosystem: Donut chart visualizing vendor distribution and type占比 of third-party models

DeepExtend¶

Uses structured prompt templates to guide LLM behavior, making it ideal for business logic alignment and building repeatable AI tasks. It primarily supports reasoning for both text and images.

DeepText¶

Compared to traditional prompt workbenches, DeepText offers deeper integration capabilities with data and models, enabling users to:

Directly embed and reference document-based knowledge in prompts
Perform inference directly on adapters or PEFT checkpoints from the training phase
Accelerate evaluation cycles by simulating production behavior using partial training results

DeepImage¶

Compared to traditional image generation tools, DeepImage provides more flexible and powerful image synthesis capabilities, allowing users to:

Generate images in any aspect ratio and size to fit diverse application scenarios
Produce multiple images in a single operation to enhance creative efficiency
Achieve creative editing and style transfer through image-to-image generation
Exercise fine-grained control over generation parameters for customized results
Utilize pre-trained models and custom checkpoints to accomplish specialized visual tasks

Dataset Management¶

Upload, version, label, and organize datasets. Supports multiple formats like JSONL, and JSON.

Knowledge Base (RAG)¶

Embed large-scale documents for retrieval-augmented generation (RAG) use cases. Supports knowledge base creation and indexing.

Model Training¶

Trigger training jobs using base models, fine-tuning methods (e.g., SFT, GRPO), with or without parameter-efficient techniques like LoRA, and customizable hyperparameters. All training tasks are fully UI-driven — no coding required.

Additionally, this section also supports the configuration of model training methods and the setup of model training parameter files.

Highlights include:

"Copy Train" button: Instantly duplicate a previous training job, preserving all configurations. Change only a few parameters (e.g., base model, dataset, LoRA rank, learning rate) to create a comparable variation.
Real-time training logs and evaluation data: Loss curves, reward points, and key metrics are fully visualized, making model progress transparent.
Multi-training comparison: Select multiple training runs and compare them side by side with one click — including performance, configuration, input and output differences.

Model Assessment¶

DeepExtension provides a powerful, batch-oriented model evaluation framework designed to help you compare model outputs at scale using real datasets.

Key capabilities include:

Assessment is based on questions sampled directly from your datasets
Four flexible evaluation modes
Each assessment job includes a preview stage to verify setup (model, dataset, measurement prompt, etc.)
Execution is done in background mode, allowing large-scale evaluations
Results are fully viewable in the UI and can also be exported as local files
Any model can be used — including third-party APIs, local adapters at any training stage, or fully deployed in-house models

This system helps teams confidently compare fine-tuned models, judge alignment quality, and standardize LLM output evaluation.

Model Management¶

Track every model in your organization through the following subcategories:

Third-party Models: External models you’ve linked (e.g., OpenAI, Anthropic, ModelScope, etc.)
Base Models: Pretrained foundation models available for use or fine-tuning
Customized Models: Outputs from the training process, typically PEFT adapters or checkpoints that are linked to a base model (also called as training artifact). These represent models in their intermediate form before merging.
Complete Models: Fully materialized models created by merging trained adapters into base models. These are independent and versioned snapshots, ready for deployment or further experimentation.
Live Models: Fully merged models that are actively served via API or integrated into internal tools. Once deployed, they function identically to base models and are ready for real-time inference or application integration.

Each model type reflects a stage in the training lifecycle — from raw training output to production-grade deployment.
The UI provides one-click actions to move models forward through this lifecycle: - Convert a Customized Models into a Complete Models - Deploy a Customized Models as a Live Models - Deploy a Complete Models as a Live Models

Delete models at any stage if they’re no longer needed

This structure gives users full control over versioning, promotion, and cleanup, ensuring that only valuable models progress while maintaining a clear and auditable lifecycle.

Settings¶

Compute Environment Configuration: Configure and define the Python environment.
Deployment Tool Configuration: Maintain integration settings for deployment tools.
Currently, only Ollama is supported. When deploying a Complete Models, the system uses this configuration to connect with your local Ollama environment. More local LLM serving tools such as LM Studio are planned for future support.
User Management: Manage team members, permissions, and access control across projects

Who Should Use This¶

This guide is ideal for:

Domain Experts who want to train and evaluate models without coding
Data Scientists / ML Engineers configuring custom training and evaluation flows
Project Managers overseeing AI workflows, security, and deployment status

Next Steps¶

You can continue with the following chapters for more detail:

DeepExtension — A unified workspace for real-world AI workflows