DNA LLM - Helix 1

DNA's are built on a proprietary in-house Large Language Model (LLM), called Helix-1. Existing LLMs demonstrate proficiency in a variety of downstream tasks. However publicly available chat models while adept in sustaining conversations, fall short in evolving dialogues naturally with end-users.

These models lack critical elements such as character depth, syntax nuances, and personality evolution. This deficiency hinders their ability to establish a conversation that sounds natural and builds a relationship beyond the capabilities of current open or closed-source chat models.

Our ambition with DNA X is to create ‘companions for life’ that not only sound natural but also evolve maintaining consistent personalities and deep knowledge about specific characters. A significant limitation of existing models is their lack of up-to-date information on notable characters and personalities which invariably impacts the quality of interactions.

In response to these challenges, we have developed an in-house Large Language Model trained to excel in vital downstream tasks for creating the companions we envision. Our model continuously integrates knowledge through an online training mechanism ensuring comprehensive character understanding without over-reliance on external data sources or techniques like retrieval augmented generation.

In our LLM, Helix-1, we exploit the “intrinsic dimension” conditioning and design multiple adapters for different tasks. For instance, one set of adapter matrices focuses on style, while another set incorporates essential knowledge about characters. To evaluate our model, we use a variety of publicly available datasets and collect multi-source character-specific data. This helps us design a diverse dataset essential for mapping the overall personality of characters, including their writing styles.

Externally, we maintain a knowledge base through retrieval augmented generation, managing it via our Vector Database tailored to our LLM.

Rather than depending on external models, we use our LLM for semantic similarity searches within our vector database. This ensures that the model’s embeddings align with our specific tasks, outperforming publicly available benchmarks.

In our initial performance evaluations, we utilized multiple benchmarks and tasks, including the Vicuna benchmark and custom criteria essential for learning personality traits. Datasets like Self-Instruct Longform and Unnatural Instruct were also used to assess our model’s capabilities. Further enhancing our model’s capabilities, we integrate feedback mechanisms to refine the interaction quality continually. This includes analyzing user interactions to understand and improve response relevance and personality consistency.