LLaMA Pro: Progressive LLaMA with Block Expansion (Paper Explained)

1. LLaMA Pro introduces block expansion for continual learning.

🥇92 00:00

LLaMA Pro enhances a large language model by adding layers to facilitate continual learning without forgetting previous knowledge.

• The method allows the model to pick up new knowledge without losing what it has already learned.
• This approach addresses the catastrophic forgetting problem associated with continual learning.

2. Post pre-training method prevents catastrophic forgetting in LLm models.

🥈85 05:30

The post pre-training method effectively and efficiently improves the model's knowledge without causing catastrophic forgetting.

• This method experiments on a corpus of code and math, yielding a new model excelling in general tasks, programming, and mathematics.
• It provides a way to retain old knowledge while adding new knowledge to the model.

3. Challenges in retaining old and adapting to new knowledge in LLm models.

🥉78 12:27

Training with only new data may distort old parameters, hindering the model's ability to adapt to new parameters.

• The model's ability to adapt to new parameters without distorting old ones depends on significant overlap between old and new data.
• Mixing old and new data during training could facilitate more drastic domain adaptations.

4. Block expansion improves model performance in specific tasks.

🥈88 15:00

Expanding blocks in LLaMA Pro enables the model to excel in new tasks such as coding, math benchmarks, and language understanding.

• The model's performance is enhanced in domains where it previously fell short.
• The data put into the model reflects the tasks it can eventually perform.

5. Understanding Progressive LLaMA with Block Expansion

🥇92 00:00

The paper explains the concept of expanding LLaMA's domain of knowledge into new areas without weakening its existing capabilities.

• The expansion involves adding identity blocks after each block in the original model.
• The experiments show that the expanded model performs as well as or better than the old models in both new and old domains.

6. Challenges in Model Retention and Parameter Divergence

🥈88 30:40

The paper raises questions about how far the new data set can be from the original training data without compromising the model's original abilities.

• It's unclear how the model retains its original abilities when trained on new tasks further apart from the original training data.
• The paper suggests the need for a good empirical evaluation to address these concerns.