IC-1519Topic and style biases in LLaMA 2 7B are learned like simple features (rapidly, with minimal capacity, concentrated at the first few tokens, magnified by learning rate) while factual knowledge is learned like complex features (slowly, requiring significant capacity, uniformly across positions, unaffected by learning rate)
Using LLaMA 2 7B finetuned on PubMed, the authors dissect the learning dynamics of the three text factors. Topic/style biases are strongest at token positions ≤10 and are acquired within 1-10k documents; at positions ≥100 they are much weaker but grow steadily. With LoRA rank 1 (0.02% of parameters), topic/style biases are learned comparably to full finetuning, whereas factual learning is significantly hindered at large training sizes. Increasing the learning rate magnifies topic/style biases in a non-saturating way, while factual learning remains consistent. Mixing Wikipedia text into the training data reduces the biases only modestly. Topic and style biases are also learned independently of each other, as shown by similar likelihood-ratio trajectories under different conditioning styles/topics.
Evidence
correlational
Key metric
Topic/style bias strongest at position ≤10, learned with 1-10k documents; LoRA r=1 (0.02% tunable parameters) recovers bias comparable to full finetuning while factual learning is significantly hindered; bias magnifies non-saturatingly with learning rate (Figure 7); mixing general text reduces bias only modestly (Figure 8)
Caveat
Capacity experiments use LoRA rank as a proxy for available capacity; the position analysis treats each document as a separate sequence (not concatenated), which differs from standard pre-training; results limited to at most 1 million documents.