Anthropic Research: Scaling Laws and Interpretability of Learning from Repeated Data

Anthropic research reveals that repeating a small fraction of training data can lead to disproportionately large performance degradations in large language models (LLMs), primarily because the model shifts from generalization to memorization. This effect is most severe when a predictable range of repetition frequency causes the model to consume a large fraction of its capacity to memorize the repeated data, effectively reducing the functional capacity of the model.

Performance Degradation and the Double Descent Phenomenon

Training LLMs on repeated data can cause a significant increase in test loss midway through training, a phenomenon known as double descent. This degradation is not linear; there is a specific range of repetition frequency where performance suffers most severely.

For example, an 800M parameter model can have its performance degraded to that of a 400M parameter model if only 0.1% of the training data is repeated 100 times, even while the remaining 90% of the training tokens remain unique. This demonstrates that repeating a relatively small amount of data can have an outsized negative impact on the model's overall capabilities.

The Mechanism of Memorization vs. Generalization

Anthropic researchers hypothesize that the peak of performance degradation occurs when the model enters a range where it can successfully memorize the repeated data. When the model allocates a significant portion of its capacity to store these memorized sequences, it loses the ability to generalize across the rest of the dataset.

This shift from generalization to memorization is linked to the detailed computations performed by the model. The research connects these observations to mechanistic interpretability, showing that data repetition disproportionately damages internal structures associated with generalization, specifically copying and induction heads.

Implications for Data Deduplication

The findings suggest that data deduplication is critical for LLM training. Whether repeated data is repeated intentionally (to upweight high-quality data) or unintentionally (due to imperfect deduplication at the sentence, paragraph, or document level), the risk of memorization over generalization is remains a high risk for model performance. The study provides a mechanistic explanation for why the model's capacity is consumed by memorization, damaging the internal structures necessary for the general intelligence of the model.