Learn Before
SFT's Reliance on Labeled Data
Supervised Fine-Tuning (SFT) is fundamentally different from pre-training due to its requirement for labeled data, whereas pre-training utilizes raw text that is readily and widely available. This dependency introduces significant challenges, as the tasks of data annotation and selection are complex, similar to other supervised machine learning domains.
0
1
Tags
Ch.4 Alignment - Foundations of Large Language Models
Foundations of Large Language Models
Foundations of Large Language Models Course
Computing Sciences
Learn After
A research lab has a powerful, general-purpose language model that was trained on a vast, unlabeled corpus of internet text. They now want to adapt this model to perform a specialized task: accurately summarizing legal documents. Based on the typical data requirements for this adaptation process, what is the most significant and immediate challenge the lab will face?
Fine-Tuning for a Niche Domain
Data Paradigms: Pre-training vs. Supervised Fine-Tuning