LianjiaTech/BELLE
BELLE
BELLE is a specialized implementation of Chinese conversational large language models, encompassing a full instruction tuning framework. It provides a pipeline for training, evaluating, and deploying models optimized for natural language understanding and dialogue tasks in the Chinese language.
The project is distinguished by its integrated approach to model refinement, combining the curation of multi-million entry instruction datasets with a distributed training pipeline. This pipeline supports both full fine-tuning and low-rank adaptation to optimize conversational performance.
The system includes a comprehensive evaluation suite that utilizes categorized test benchmarks and automated scoring prompts to assess output quality. For deployment, it provides a quantized runtime that enables these models to run locally and offline on both desktop and mobile devices.
Features
- Chinese Conversational LLMs - Implements a specialized conversational large language model optimized for natural language understanding in Chinese.
- Automated Output Evaluation - Uses secondary large language models to score outputs based on predefined rubrics and test datasets.
- Training Pipelines - Implements a full pipeline for creating and fine-tuning large language models optimized for the Chinese language.
- Instruction - Provides a system for gathering and curating multi-million entry Chinese language instruction datasets.
- Distributed Gradient Synchronization - Implements mechanisms to coordinate weight updates across multiple GPU nodes for large-scale parallel training.
- Instruction Tuning - Utilizes curated prompt-response pairs to refine base model behavior for improved conversational alignment.
- Instruction Tuning Frameworks - Offers an integrated system for curating datasets, fine-tuning, and evaluating models for conversational performance.
- LoRA Training - Employs Low-Rank Adaptation to reduce memory requirements by updating only a small subset of weight matrices.
- Training Pipelines - Provides an end-to-end workflow combining full fine-tuning and low-rank adaptation to adapt models to instruction sets.
- Distributed Fine-Tuning - Implements distributed training systems to support both full fine-tuning and low-rank adaptation.
- Model Evaluation Suites - Provides a comprehensive suite of categorized test benchmarks and scoring prompts to assess model outputs.
- Weight Quantization - Converts high-precision floating point weights into lower-bit integers to decrease model size and memory usage.
- LLM Evaluation - Provides tools for measuring output quality and accuracy using standardized test sets and automated scoring.
- Local Chat Applications - Provides a cross-platform application for deploying quantized models on mobile and desktop for offline interaction.
- Offline Deployments - Enables the execution of quantized models on local devices for private interaction without internet access.
- Model Inference Wrappers - Ships a shared interface that allows the model engine to run on both mobile and desktop operating systems.
- Quantized Inference Runtimes - Provides a cross-platform runtime environment designed to execute compressed and quantized language models.
- Data Expansion - Explores the impact of instruction data scaling on model performance.
- Foundation Models - Instruction-tuned language models based on LLaMA and Alpaca architectures.
- Generative Language Models - Open-source instruction-tuned model for Chinese language applications.
- Instruction Tuning Datasets - Large-scale Chinese instruction-tuning datasets for model training.
- Language Models - A project focused on developing Chinese-language conversational models.
- Open Source Models - Provides Chinese-optimized conversational models.
- Specialized Domain Models - Instruction-tuned model for general and vertical tasks.
- Text LLM Models - Instruction-tuned models based on BLOOM and LLaMA architectures.
- Instruction Datasets - Self-instruct generated dataset for diverse task training.
