Nexusflow has recently launched Athene-Llama3-70B, a high-performance open-weight chat model that’s been fine-tuned from Meta AI’s earlier model, Llama-3-70B. The improvement in terms of performance is quite significant with the new model achieving an impressive Arena-Hard-Auto score of 77.8%, surpassing models like GPT-4o and Claude-3.5-Sonnet. This is a substantial improvement from Llama-3-70B-Instruct, the predecessor which scored only 46.6%.
The considerable difference in performance can be attributed to Nexusflow’s specialized post-training process. This was developed in a bid to bolster specific behaviors in the model. To tap into the full potential of the Llama-3-70B model, Nexusflow created several internal benchmarks. These benchmarks were used to measure Latent Language Models (LLMs) abilities toward various tasks such as instruction following, coding, creative writing, and multilingual tasks.
The assessment results then served as the basis for curating high-quality preference data. This data was targeted for training reinforcement through human feedback, better known as RLHF. This approach effectively resulted in marked performance improvements compared to the predecessor model, enhancing key aspects such as precision in instruction following, math and reasoning, comprehensive coding assistance, inspired creative writing, and multilingual mastery.
Athene-70B serves as a testament to Nexusflow’s capabilities in customizing models through targeted post-training to meet specific enterprise needs. This accomplishment builds on Nexusflow’s previous successful endeavours with the Starling-7B and NexusRaven-V2 models. Nexusflow is intent on improving these models further to meet the standards of enterprise-grade applications.
Nexusflow offers bespoke solutions to assist businesses excel in General AI (GenAI) co-pilot and agent technologies. The company encourages organizations keen on enhancing their AI initiatives to explore how Athene-70B might be of use. To further facilitate such exploration, businesses are invited to contact Nexusflow for additional information and potential collaboration opportunities.
Nexusflow has also developed an Athene-Llama3-70B Model Card which credit for goes to project researchers. The company encourages the public stay updated with their latest work through their Twitter and LinkedIn platforms and joining their Telegram Channel. They also run a 46k+ ML SubReddit space and regularly update their newsletter. Moreover, Nexusflow is actively involved in hosting AI webinars to discuss new updates and future projects.
