Sarus SarusLLM

SarusLLM is a privacy-preserving LLM fine-tuning platform designed for businesses and developers who want to use open-source large language models (LLMs) on sensitive data without risking data leakage or memorization of private information. The platform addresses a core risk with LLM fine-tuning: models trained on sensitive data tend to memorize and regurgitate that information when prompted. SarusLLM mitigates this through two primary mechanisms: 1. **Data Clean Rooms for LLM Training:** Data scientists can explore, preprocess, and feed data into LLMs without directly accessing the underlying data. Only high-quality synthetic data and differentially-private statistics can be retrieved from the clean room. Standard AI and GenAI tooling is wrapped via the Sarus Python SDK. 2. **Differentially-Private Fine-Tuning:** Differential Privacy (DP) guarantees are embedded into the LLM fine-tuning process using Differentially-Private Stochastic Gradient Descent (DP-SGD). This is activated via a single fit parameter, ensuring no personal data is embedded in the fine-tuned model weights. The platform is deployed within the user's own infrastructure by installing the Sarus application and using the Sarus Python SDK to launch fine-tuning jobs. GPU orchestration is handled automatically. Fine-tuned models can then be deployed to production as normal. SarusLLM supports fine-tuning of open-source LLM families (e.g., Mistral, LLaMA, GPT-2) across various infrastructures and data types. It targets use cases in healthcare, banking, insurance, retail, and marketing, including synthetic data generation, fraud detection, and RAG-based assistant development.