Glossary

The glossary provides you with core terms covering AI technologies and related fields, serving as a foundational reference for understanding product features and technical principles.

Term Category	Term Name/Abbreviation	Definition
Model & Algorithm Technology	LoRA（Low-Rank Adaptation）	A lightweight model fine-tuning technique that reduces fine-tuning parameters through low-rank matrix factorization, applicable to text-to-image, large language models, and other scenarios.
	LLM（Large Language Model）	Large-scale language models trained on massive text data, supporting tasks such as text generation and question answering, e.g., Qwen2.5, GLM-4.
	SFT（Supervised Fine-Tuning）	Optimizing pre-trained models using labeled data to improve performance on specific tasks, suitable for small dataset scenarios.
	DPO（Direct Preference Optimization）	Optimizing models based on human preference data to enhance the conversational quality of large language models.
	Transformers	An architecture based on self-attention mechanisms, used for NLP tasks and one of the core architectures of LLMs.
	Diffusers	A generative image model library supporting text-to-image and image-to-image generation, such as Stable Diffusion.
	Function Calling/Tool Calling	The capability of large models to invoke preset functions/tools, enabling cross-tool collaboration (e.g., calling a weather API).
Data & Training Related	Dataset	A structured collection of data used for model training, divided into public and private, supporting multiple formats.
	Dataset Card	A file (README.md) that records dataset metadata (license, language, etc.) to help understand the dataset.
	Model Card	A document describing model information, including use cases and limitations, hosted as the README.md in the model repository.
	Token	The basic unit for models to process text, used for calculating length and billing.
API & Service Technology	Serverless API	API services that eliminate the need for server management, billed by the number of calls, supporting functions such as text generation.
	API Workflow	A visual workflow engine that connects multiple model APIs to automate tasks (e.g., "text→image" generation).
	Access Token	Identity credentials for API calls, divided into full-resource and partial authorization, with permission controlled by binding resource packages.
	Model Context Protocol (MCP)	A protocol that defines how models request tools to perform operations, interact with the environment, and obtain results from external systems.
	Temporary Token	Short-term valid tokens, suitable for client-side calls with higher security.
	Failover Mechanism	Automatic switching to other compute resources when the primary resource fails, enabled via the request header X-Failover-Enabled.
Image and Multimedia Technology	Text-to-Image	Generate images based on text, such as the FLUX.1 model, which can control effects through parameters.
	Image-to-Image	Generate new images based on the original image, supporting style transfer and local repainting.
	OCR (Optical Character Recognition)	Technology for extracting text from images, supporting multi-language, table, and formula recognition.
	TTS (Text-to-Speech)	Converting text to speech with voiceprint customization, such as the ChatTTS model.
	ASR (Automatic Speech Recognition)	Converting speech to text, supporting multiple languages and dialects.
Deployment & Computing Related	Model Engine	A managed model deployment service that supports custom computing power, suitable for production environments.
	Edge Device	Edge devices (e.g., mobile phones) that support the deployment of lightweight models for low-latency inference.
	GPU	Parallel computing hardware suitable for model training and inference, such as NVIDIA A10.
	vllm	A high-performance LLM inference library that optimizes the concurrency and latency of conversational models.
	Docker	A containerization tool that packages applications and their environments, supporting custom configurations.
Parameter & Configuration Related	Temperature	Controls generation randomness (0-1); lower values result in stable output, while higher values lead to more creative results.
	Top_p	Controls diversity (0-1), used in conjunction with temperature.
	Guidance Scale	A parameter in image generation that controls the influence of text prompts.
	JSON Schema	A specification that constrains the output format of models to ensure structured data generation.