finetuned models

Fine-tuned models: Specialists created by adapting a broad pre-trained foundation model (e.g., Llama 3, GPT-4) on a smaller, high-quality, domain-specific dataset.

Fine-tuning is the critical process of taking a general-purpose pre-trained model and further training it on a task-specific dataset: this is transfer learning in action. The goal is specialization: adapting the model's weights to a specific domain (e.g., legal, medical, finance) to improve accuracy and relevance, not to inject new general knowledge. Techniques like Parameter-Efficient Fine-Tuning (PEFT), specifically LoRA, allow for this adaptation with minimal computational overhead (e.g., using a single GPU) and prevent catastrophic forgetting of the base model's capabilities. A fine-tuned, smaller model (e.g., Llama 3 8B) can often outperform a much larger, general model on its targeted task, reducing inference latency and cost in production environments.

https://huggingface.co/docs/peft/en/index