Perplexity adopted NVIDIA NeMo, leveraging its reliability, flexibility, and ease of use to create custom models for their online answer engine. They utilized several data processing and advanced model alignment techniques supported by NeMo:
- Supervised Fine-Tuning: NeMo's capabilities in handling data distributed across multiple nodes enabled Perplexity to scale their training processes efficiently.
- Direct Preference Optimization (DPO): This allowed Perplexity to enhance the performance of pretrained models to align with human preferences, tailoring the models to users’ needs.
- Proximal Policy Optimization (PPO): This alignment technique improved the outcomes of training models for complex tasks, such as playing games and controlling robots, with improved outcomes.
Within a few days of a new open-source release, the team had a new Sonar model that was 20% improved over the base model on search.
Perplexity has applied fine-tuning to frontier models, including Llama and Mistral model families, and is leveraging retrieval-augmented generation to deliver precise, concise answers based on the retrieved data. This level of customization has enabled Perplexity to achieve high accuracy and relevance in their AI applications.
Additionally, NeMo's ease of use, breath of supported model architectures, and high training throughput allowed Perplexity to quickly experiment and find the best-tuned models for their applications.