Show HN: Chorus –Let users improve the LLM without their data leaving the device (github.com)

Chorus, a newly announced framework, enables federated fine-tuning of large language models (LLMs) while ensuring that users' data remains on their devices. Utilizing LoRA (Low-Rank Adaptation) adapters, Chorus allows multiple clients to train on their private datasets and send adapter updates (deltas) to a central server. This approach replaces traditional methods that often inadequately average model weights across clients with a more precise technique called FedEx-LoRA. This method tackles the challenges of mathematically exact aggregation by leveraging Singular Value Decomposition (SVD) to accurately compute the global model updates without requiring raw data exchange. The significance of Chorus lies in its ability to enhance privacy in AI/ML applications while still enabling collaborative model improvements. By maintaining data locality and applying innovative aggregation techniques, it promises to mitigate risks associated with data leaks and non-compliance with data regulations. The framework includes built-in security features, such as differential privacy and Byzantine defenses, further bolstering its appeal for sensitive applications. As the AI/ML community increasingly prioritizes data privacy, Chorus represents a crucial development toward creating collaborative yet secure AI systems.