Data Darwinism Part I: Unlocking the Value of Scientific Data for Pre-training

TL;DR

This paper introduces Data Darwinism, a taxonomy for data-model co-evolution, and demonstrates how advanced data processing techniques improve scientific data pre-training, leading to significant performance gains on benchmarks.

Contribution

It proposes a ten-level taxonomy for data quality and processing, and validates its effectiveness through constructing and pre-training models on the Darwin-Science corpus, achieving state-of-the-art results.

Findings

Darwin-Science outperforms baselines by +2.12 and +2.95 points on benchmarks.

Higher-level data processing (L5) yields +1.36 total gain.

Systematic data refinement unlocks latent data value.

Abstract

Data quality determines foundation model performance, yet systematic processing frameworks are lacking. We introduce Data Darwinism, a ten-level taxonomy (L0-L9) that conceptualizes data-model co-evolution: advanced models produce superior data for next-generation systems. We validate this on scientific literature by constructing Darwin-Science, a 900B-token corpus (L0-L5). We identify a learnability gap in raw scientific text, which we bridge via L4 (Generative Refinement) and L5 (Cognitive Completion) using frontier LLMs to explicate reasoning and terminology. To ensure rigorous attribution, we pre-trained daVinci-origin-3B/7B models from scratch, excluding scientific content to create contamination-free baselines. After 600B tokens of continued pre-training, Darwin-Science outperforms baselines by +2.12 (3B) and +2.95 (7B) points across 20+ benchmarks, rising to +5.60 and +8.40 points on domain-aligned tasks. Systematic progression to L5 yields a +1.36 total gain, confirming that higher-level processing unlocks latent data value. We release the Darwin-Science corpus and daVinci-origin models to enable principled, co-evolutionary development.