What technique involves freezing the original model parameters?

The technique that involves freezing the original model parameters is known as LoRa (Low-Rank Adaptation). In this approach, rather than updating all the parameters of a pre-trained model, only a small set of additional low-rank parameters is trained while keeping the rest of the model parameters fixed. This allows for efficient fine-tuning of models with significantly less computational resources, as the original model weights remain unchanged.

LoRa is particularly useful in scenarios where you want to adapt a large, pre-trained model to specific tasks without the need for extensive retraining, which can be both time-consuming and resource-intensive. This method helps in leveraging the knowledge captured in the original model while allowing for targeted improvements based on the new data.

The other techniques mentioned involve different methods. For example, DPO (Direct Preference Optimization) focuses on optimizing preferences in generated outputs rather than freezing parameters. Prompt Learning emphasizes designing effective prompts to elicit desired outputs from models instead of modifying their internal architecture. Adam is an optimization algorithm that updates model parameters during training based on the gradients, without any freezing of weights. Thus, LoRa stands out as the specific technique that entails freezing model parameters to enhance efficiency during the fine-tuning process.