Persimmon
About
The Persimmon family of large language models (LLMs) by Adept AI features decoder-only transformer models that excel despite their relatively small parameter counts. Persimmon-8B, the most renowned model in this family, offers an impressive context size of 16K tokens, enabling it to manage longer inputs and preserve more context during text generation 12. Adept AI emphasizes practical evaluation of the models, focusing on direct text generation over implicit probabilities 1. They are distributed under an Apache license to encourage community contributions and further development 12. Although the base model's performance is similar to Llama 2 with less training data, its instruction-tuned variant, Persimmon-8B-FT, outperforms in various benchmarks 1. The architecture incorporates enhancements like squared ReLU activation and query/key layernorm for increased efficiency, and Adept provides fast inference code, blending C++ speed with Python's flexibility 1.