Tenyx Launches Fine-tuning Platform to Fix Catastrophic Forgetting in Large Language Models

Tenyx, a leader in voice AI systems that automate customer service functions for the enterprise, today announces a novel solution to one of the most significant challenges in AI: catastrophic forgetting during fine-tuning of large language models (LLMs). With this groundbreaking methodology, Tenyx helps businesses adapt LLMs to their unique requirements without compromising foundational knowledge and protective safeguards. To get started or to sign up for access, visit here: https://www.tenyx.com/fine-tuning

“In the rapidly evolving landscape of AI, our commitment has always been to address its inherent challenges head-on. With this novel methodology, we’re not just pioneering an advanced solution; we’re revolutionizing the way enterprises utilize LLMs. Our innovation ensures that businesses no longer have to choose between customization and core capabilities. They can confidently enjoy the best of both worlds,” said Itamar Arel, CEO and founder of Tenyx.

Drawbacks of Traditional Fine-tuning Techniques

The conventional approach to fine-tuning LLMs poses inherent risks. Training models with new data to perform better in certain areas can cause unintentional loss or degradation of previously learned capabilities. The complexity of these models makes it exceedingly challenging to pinpoint and rectify these distortions. Current fine-tuning solutions rely primarily on Low-Rank Adaptation, or LoRA, a technique that lacks the ability to mitigate forgetting effects. Additionally, conventional schemes used for fine-tuning risk eroding the safety measures established by RLHF (reinforcement learning from human feedback). This mechanism, vital for preventing harmful model outputs, can be inadvertently weakened or retracted during fine-tuning using traditional methods.

Fixing Catastrophic Forgetting in LLMs

By leveraging a novel mathematical interpretation of the geometric representations formed during the initial LLM training, Tenyx’s methodology alleviates the aforementioned drawbacks and ensures that models can be customized to a specific customer domain without significant loss of prior capabilities. This approach not only improves the retention of prior knowledge and reasoning abilities, but also retains the RLHF protection, providing an unparalleled boost in enterprise use of LLMs. Moreover, safer fine-tuning is aligned with changes to the regulatory environment, specifically as they relate to the recent White House executive order on Safe, Secure, and Trustworthy AI.

By evaluating popular enterprise and open-source finetuning algorithms, Tenyx’s pilot study boasts higher proficiency, remains safer, and better retains initial knowledge in three key ways:

• Safety: For all models, safety capabilities are drastically reduced after fine-tuning. Yet, Tenyx fine-tuning witnessed an 11% reduction, compared to OpenAI’s -66%, Together AI’s -94%, and LoRA’s -91%

• Proficiency: despite OpenAI’s GPT 3.5 Turbo being more initially proficient due to the model having more parameters, Tenyx’s Llama-2 7B was the most proficient after fine-tuning.

• Knowledge: Similar to safety, knowledge retention drastically reduces after fine-tuning. However, Tenyx mitigated catastrophic forgetting the best with a 3% loss, compared to OpenAI’s 10%, Together AI’s 40%, and LoRA’s 43%.

“We’ve worked with a wide variety of customers to help them tackle the challenges in achieving optimal fine-tuning results,” said Jaron Waldman, Chief Product Officer at Cohere. “Tenyx is taking an innovative approach aimed at enhancing overall quality and reliability, while ensuring that LLM capabilities are not negatively impacted.”

“Catastrophic forgetting is a well-known problem in deep learning, which still affects even large, capable models. When trained on data from a new domain, models generally perform better on that domain while unintentionally modifying earlier capabilities. Tenyx has a strong research team who are exploring important new ideas for addressing this difficult challenge,” said Noah Goodman, Associate Professor at Stanford University.

About Tenyx

Tenyx is a privately held company in Los Altos, California delivering intelligent voice AI agents for customer service automation. Leveraging proprietary AI technology, the company is creating integratable solutions that are naturally engaging, capable of understanding context and able to offer the sense of immediacy found in human interactions. Tenyx helps businesses and their employees minimize wait times, boost service productivity and improve customer experience quality.

Contact Information

Eric Gonzalez

Media Contact

VSC for Tenyx

tenyx@vsc.co