Athene
About
The Athene family of LLMs originates from Nexusflow, a company that excels in post-training optimization of large language models. Its flagship model, Athene-Llama3-70B (commonly known as Athene-70B), is an open-weights model fine-tuned from Meta AI's Llama-3-70B-Instruct through reinforcement learning from human feedback (RLHF) 5. This meticulous post-training approach has notably enhanced its capabilities, leading to an impressive 77.8% score on Arena-Hard-Auto, a benchmark that strongly aligns with human evaluation on Chatbot Arena 5. As a result, Athene-70B is among the top-performing open-source models, on par with leading proprietary models 5. Nexusflow's post-training efforts mainly targeted improving the model’s instruction following, reasoning, coding, creative writing, and multilingual abilities 5. Additionally, the Athene series includes smaller models and others hinted at as Athene v2, v3, and v4 in different references 3811, though these are less detailed in the provided information.