Challenge of Opaque Pre-Training Data in Fine-Tuning
The lack of transparency regarding the specific data used during a model's pre-training phase creates a significant challenge for fine-tuning. Without knowing which instruction-response patterns the model has already been exposed to, it becomes difficult to determine which mappings need to be explicitly taught during the fine-tuning stage.
0
1
Tags
Ch.4 Alignment - Foundations of Large Language Models
Foundations of Large Language Models
Foundations of Large Language Models Course
Computing Sciences
Related
Fine-Tuning Pre-trained Models for Downstream Tasks
Instruction Fine-Tuning
Superficial Alignment Hypothesis
Challenge of Opaque Pre-Training Data in Fine-Tuning
A team develops a large language model pre-trained on a massive, diverse corpus of text from the internet. When initially tested on the task of generating concise summaries of legal documents, its performance is poor and unstructured. The team then collects a small, curated dataset of 500 legal documents and their corresponding expert-written summaries. After training the model on this small dataset, its ability to summarize new legal documents improves dramatically. Which statement best analyzes the role of this second training phase?
Critiquing a Model Training Hypothesis
Implicit Learning of Instruction-Response Mappings During Pre-training
Explaining the Impact of Targeted Training
Challenge of Opaque Pre-Training Data in Fine-Tuning
A machine learning engineer claims, "A language model's ability to follow instructions is exclusively a result of the targeted examples shown during its fine-tuning stage. The pre-training phase only provides it with general world knowledge and language structure."
Which of the following statements provides the most accurate evaluation of this claim?
Explaining Unexpected Model Capabilities
Explaining Emergent Zero-Shot Abilities
Learn After
Primary Source of Out-of-Distribution Generalization: Pre-training vs. Fine-tuning
Diagnosing Inconsistent Fine-Tuning Performance
A development team is fine-tuning a large, pre-trained language model to act as a specialized legal assistant. They notice that the model quickly masters tasks related to contract law after seeing only a few examples, but struggles to generate accurate summaries of intellectual property case law, even with a large number of fine-tuning examples. What is the most likely underlying reason for this discrepancy?
The 'Unknown Unknowns' of Fine-Tuning Strategy