My finetuned models beat OpenAI’s GPT-4

The post outlines the process and results of finetuning models to perform structured data extraction from press releases. The author reports that their finetuned models outperform OpenAI's GPT-4 models in terms of accuracy on several metrics. The process, however, involved considerable complexity and performance trade-offs. The data was stored and evaluated using the Hugging Face Hub, and the tests revealed that finetuned models can offer advantages in terms of accuracy, cost, data privacy, and control. The author also discussed the challenges in managing multiple models and the need for efficient evaluation systems.