Less In, More Retained: Improving Factual Memorization through Training Data Pruning

LLMs struggle to memorize factual knowledge within their parameters, which often leads to hallucination and performance degradation. Research shows that factual accuracy deteriorates when the amount of information in the training data exceeds the model's Capacity, or when the frequency distribution of facts follows a power law and is imbalanced.

To address this, a Data Selection technique is proposed that uses only Training Loss to limit the amount of data and flatten the frequency distribution.

Experimental results show that this methodology achieved the following outcomes:

• On datasets containing high-entropy facts, it raised factual accuracy up to the model's capacity limit.
• When trained on the Wikipedia corpus, the GPT2-Small (110M) model memorized 1.3x more entity facts than with the conventional approach.
• This result matches the performance of a 1.3B-scale model trained on the full dataset.