Core Concepts
Developing Polaris, a safety-focused Large Language Model (LLM) constellation for real-time patient-AI healthcare conversations, with specialized agents and rigorous training protocols.
Abstract
The content introduces Polaris, a unique LLM constellation system designed for real-time patient-AI healthcare conversations. It outlines the architecture, training protocols, data sources, and evaluation results. The system includes a primary agent focused on engaging conversations and specialist support agents for healthcare tasks. Training involves proprietary data, simulated conversations, and clinician evaluations to ensure medical accuracy and rapport building.
Introduction to Polaris: A safety-focused LLM constellation for healthcare conversations.
System Overview: Objectives, architecture details, orchestration methods, and LLM specifics.
Conversational Alignment: Data sources, instruction tuning stages, conversation nuances.
Specialist Support Agents: Privacy & compliance specialist, medication specialist details.
Evaluation Results: Clinician evaluations showing Polaris performs comparably to human nurses.
Stats
We train our models on proprietary data, clinical care plans, healthcare regulatory documents, medical manuals.
Our models range from 50B to 100B parameters with decoder-only transformer-based architecture.
Training runs were performed using Nvidia H100 GPUs with DeepSpeed.
During inference we deploy some models in bf16 precision and others in int8 precision.
Quotes
"We develop Polaris 2...for real-time patient-AI healthcare conversations."
"Our one-trillion parameter constellation system is composed of several multi-billion parameter LLMs as cooperative agents."
"We present the first comprehensive clinician evaluation of an LLM system for healthcare."