Enhanced base LLMs: Perplexity Sonar+DeepSeek R1 optimized by fine tuning Llama

FounderCulture, Storytell and UpHonest Capital have teamed up to create a Meetup series called "Building The Future of AI."

🟢 This event will be live in San Mateo, CA -- also be live-streamed on this Zoom link if you can't join in person. The Zoom will start at 6:30pm PT.

Join us for a robust discussion led by Tim Kellogg and moderated by DROdio.

Enhanced base LLMs: Perplexity Sonar+DeepSeek R1 optimized by fine tuning Llama

The rapid evolution of large language models has brought fine-tuning to the forefront as a critical strategy for enhancing performance. This session will explore the cutting-edge approaches used to optimize two leading models: Sonar and DeepSeek R1, both fine-tuned on Llama architectures.

• Perplexity's Sonar, built on Llama 3.3 70B, has been further trained to excel in factuality and readability, making it ideal for search applications. Its integration with Cerebras inference technology enables blazing-fast token processing, allowing for real-time responses. The model’s open-source nature and Mixture of Experts (MoE) architecture, which activates only a subset of parameters per query, will be discussed in depth.
• DeepSeek R1, a fine-tuned version of Llama 70B using samples generated by DeepSeek R1, has demonstrated exceptional performance on tasks like MATH-500 (94.5%) and AIME 2024 (86.7%). It supports a full 128k context window and optimized token throughput, significantly enhancing its reasoning capabilities.

This meetup will delve into the methodologies behind fine-tuning, the technological advancements that support it, and the emerging evidence from models like Sonar and DeepSeek R1. Discussion will encompass techniques, infrastructure needs, and cost considerations involved in fine-tuning LLMs for reasoning tasks, revealing how businesses and researchers can leverage this practice to enhance performance and yield better results.

Join us as we explore how fine-tuning is challenging the traditional notion that base LLMs should not be fine-tuned, and how these optimized models are setting new benchmarks for efficiency and accuracy in AI search and reasoning.

🎁 FounderCulture is powered by our Patrons -- consider becoming a Patron to support our vision of improving the odds for Founders globally (you'll get early access to RSVP to our events!)