Reproduce “Data Darwinism Part II: DataEvolve -- AI can Autonomously Evolve Pretraining Data Curation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tiantian Mi, Dongming Shan, Zhen Huang, Yiwei Qin, Muhang Xie, Yuxuan Qiao, Yixiu Liu, Chenyang Zhou and 1 more
DataEvolve automates category-specific data-cleaning strategy design by scoring cleaned samples and refining instructions with diagnostic feedback. It is applied to eight academic-science categories, producing data for 3B-parameter models trained on 500B tokens. Those models score 44.13, the authors’ average across 18 benchmarks, 3.96 points above the raw-data models. Results vary by task and do not establish superiority beyond these domains.
Key insights
- Automated, evolutionary optimization of data curation strategies effectively improves data quality, leading to superior LLM performance, particularly on knowledge-intensive tasks.
- Highly effective curation strategies are primarily cleaning-focused, emphasizing targeted noise removal, format normalization, and domain-specific preservation over content transformation.
- Data cleaning differentially impacts model performance: it significantly benefits knowledge memorization tasks but can degrade performance on tasks requiring informal language understanding.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.