Reproduce “Amuro and Char: Analyzing the Relationship between Pre-Training and Fine-Tuning of Large Language Models” on OpenResearch

About this paper

Kaiser Sun, Mark Dredze

An investigation into large language model training dynamics reveals that prolonged pre-training develops latent capabilities, enhancing subsequent fine-tuning on certain tasks even without direct initial improvement, while supervised fine-tuning can induce asymmetrical forgetting, particularly for generative skills when specializing in classification.

Key insights

• Instruction-following capabilities mature during extensive pre-training, while fine-tuning primarily teaches models how to format responses for specific tasks.
• Fine-tuning can lead to asymmetrical cross-task forgetting; specialization in classification tasks significantly degrades generative abilities, but generative fine-tuning does not substantially impair classification.
• Continued pre-training imparts valuable latent information for tasks not explicitly learned during pre-training, enhancing the model's capacity for fine-tuning on those tasks, even if the base model's performance on them remains unchanged.

Summary by alphaXiv.