Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI advancements.
To fully utilize large language models (LLMs), businesses need to customize them using domain-specific data. This fine-tuning ensures the models produce relevant outputs. However, there’s a major issue with this: fine-tuning on data different from the original training set can make the model forget what it previously learned—a phenomenon known as “catastrophic forgetting.” This diminishes the model’s knowledge, reasoning abilities, performance, and overall usability.
Today, Tenyx, a voice AI agent company, is announcing a new method to fine-tune LLMs that addresses this problem. Their approach allows businesses to adapt LLMs to their specific needs without losing the foundational knowledge or safety protocols.
Catastrophic forgetting has long been a challenge in machine learning. Many in the field have assumed that continually training on new data, while integrating old data, would resolve this. However, this often leads to the loss of critical capabilities and exposes the model to harmful and biased content. Fine-tuning is becoming essential for leveraging LLMs in enterprise settings. Yet, data scientists typically don’t have access to the full training dataset, and traditional methods don’t adequately address forgetting effects, leading to potential legal liabilities and diminished model performance.
For instance, a general model like LLaMA 7B might be used as a customer service chatbot but needs fine-tuning with typical conversations to function well. Conventional techniques, such as Low-Rank Adaptation (LoRA), may improve the model’s responses to specific inputs but at the cost of losing general knowledge and reasoning skills.
LoRA is popular for its efficiency in memory and computational resources but was not designed to prevent catastrophic forgetting. Updating model weights for new data distributions often leads to unpredictable distortions. This can weaken or eliminate safety measures established through reinforcement learning from human feedback (RLHF), which is crucial for avoiding biased and harmful outputs.
Current solutions to mitigate catastrophic forgetting involve extensive manual work by machine learning engineers. This process is unreliable, inconsistent, and time-consuming, with no automated methods available to streamline it.
Tenyx’s method determines which model parameters can be updated to allow learning new data while retaining almost all previously learned information. They project updates made during fine-tuning to a space that doesn’t interfere with the pre-trained data representation. This ensures minimal to no catastrophic forgetting.
Tenyx’s platform leverages a novel mathematical interpretation of the geometric data representations within transformer networks, ensuring that updates during fine-tuning don’t disrupt the captured information. This method retains RLHF protections and complies with regulatory requirements like the White House Executive Order on Safe, Secure, and Trustworthy AI.
In pilot studies comparing popular enterprise and open-source fine-tuning algorithms, Tenyx demonstrated significant gains:
– Safety: Tenyx fine-tuning resulted in an 11% reduction in safety concerns, compared to much higher reductions in other models.
– Proficiency: Llama-2 7B, following Tenyx fine-tuning, showed the highest proficiency among tested models.
– Knowledge: Tenyx had the least loss in knowledge at 3%, significantly outperforming others.
Catastrophic forgetting remains a known issue in deep learning, even for advanced models. Training on new domain data often improves performance in that area but can unintentionally degrade earlier capabilities. Tenyx’s innovative approach offers a promising solution to this enduring challenge.