Synthetic data: save money, time and carbon with open source
Hugging Face
Researchers demonstrated that using open-source language models to generate synthetic training data for custom sentiment analysis models costs 1,127 times less than using GPT-4 while maintaining equivalent accuracy. A custom RoBERTa model trained on synthetic data analyzed financial news for $2.7 and 0.12 kg CO2 compared to $3,061 and 735-1,100 kg CO2 with GPT-4, with 0.13-second latency versus multiple seconds. This approach enables companies to build task-specific models without the expense, latency, and third-party data exposure of commercial LLM APIs.