AI-powered
podcast player
Listen to all your favourite podcasts with AI-powered features
Investing Time in Data Preparation for Model Training
Investing significant time in data work and cleaning is crucial for training embedding models. The process involves two stages: large-scale contrastive pre-training with around 240 million pairs of semantically related sentences, and smaller scale contrastive fine-tuning including hard negatives to enhance retrieval performance. The addition of hard negatives aids in pushing model performance further.