About
The Aquila 2 family is a series of bilingual large language models designed to proficiently handle Chinese and English languages. These models vary significantly in size, from 7 billion to 70 billion parameters, offering a broad range for different computational needs. They are developed using the advanced HeuriMentor (HM) framework, which enhances the training process by dynamically adjusting data distributions, thereby boosting efficiency and model performance. This framework includes components such as the Adaptive Training Engine (ATE) and Training State Monitor (TSM). The Aquila 2 models consistently perform well on key benchmarks and have been open-sourced to encourage further innovation. Notably, the Aquila2-34B model retains performance standards even when quantized to Int4 format, highlighting its efficiency and accuracy. The AquilaChat2 variants are specially fine-tuned for conversational applications, demonstrating the versatility of the Aquila 2 family 14.