81How do you do continued pretraining to adapt an LLM to a new domain without forgetting general ability?▼hardMetaDatabricksSnowflake1 replies◆ premiumContinued pretraining adds domain knowledge that fine-tuning cannot, yet done carelessly it wrecks general ability. The signal is the replay ratio, learning-rate rewarming, and how you measure forgetting.Open full answer →
123How would you adapt an LLM to a specialist domain like law or medicine, end to end?▼hardGoogleMicrosoftIBM◆ premiumThe staged recipe is the easy half. The half that sinks real projects is data licensing, an evaluation bar that needs actual domain experts, and knowing which parts of the problem fine-tuning cannot solve at all.Open full answer →
