Chinese researchers have unveiled a new LLM, the FLM-101B, a decoder-only LLM boasting a remarkable 101 billion parameters. This development provides a cost-effective alternative for both research and practical applications.
What makes FLM-101B stand out is its exceptional performance achieved on a relatively modest budget. While it’s well-known that training LLMs from scratch can require astronomical investments, the creators of FLM-101B have shown that it’s possible to train a model with 101 billion parameters using just a $100K budget.
The experimental results are nothing short of impressive. FLM-101B has demonstrated performance levels comparable to established and resource-intensive models like GPT-3 and GLM-130B. This comparison highlights the tremendous potential of this cost-effective model, particularly on IQ benchmarks with complex contexts not present in the training data.
In a move that underlines their commitment to advancing AI research and development, the creators of FLM-101B have made this model open-source. Researchers and developers worldwide can now access and leverage this 101B-scale LLM for various applications, spanning both the Chinese and English languages.
The FLM-101B model employs a unique training approach. It rapidly accumulates knowledge from a smaller 16-billion-parameter model in the initial stages of training and progressively scales up to 101 billion parameters. This incremental approach significantly reduces training costs, making it financially feasible for a broader range of projects.
One standout feature of FLM-101B is its support for efficient window size expansion during inference. This is achieved through the use of xPos rotary position embedding, allowing the model to handle a broader context, enhancing its adaptability and usability.
FLM-101B was trained on a cluster of 24 DGX-A800 GPU servers in less than 26 days. This impressive feat underscores the model’s scalability and efficient resource utilization. The model’s training codebase, adapted from Megatron-LM, will soon be available as open-source, providing valuable insights for the AI community.
The creators of FLM-101B acknowledge potential limitations, including the model’s exposure to unsafe examples in the training corpus due to the open nature of the dataset. This caveat serves as a reminder of the importance of responsible AI usage and content moderation.
While FLM-101B has achieved remarkable results, the creators acknowledge areas for improvement. The model’s inference process, while powerful, is not yet fully optimized, leading to higher resource usage and reduced speed. However, plans are underway to introduce Flash Attention in inference, addressing this limitation.
Source: mPost
