Research map / Large language models / Language, safety and interpretability
D4: Improving LLM Pretraining via Document De-Duplication and Diversification
Kushal Tirumala, Daniel Simig, Armen Aghajanyan, Ari Morcos
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply D4 selection (SemDeDup then SSL Prototypes) at R=0.25 for pretraining | OPT-6.7B · validation perplexity, 100B tokens | training efficiency: 20% fewer update steps to same perplexity | Improved |
| apply D4 selection at R=0.25 | OPT-6.7B · 16 NLP tasks, 0-shot average | accuracy: 2% improvement | Improved |
| repeat D4-selected data for 2 epochs instead of single epoch on random new data | OPT-1.3B · Instruction + Answers perplexity | perplexity: 14.19 -> 13.85 | Improved |
| repeat randomly selected data for 2 epochs instead of single epoch on all new data | OPT-1.3B · Instruction + Answers perplexity | perplexity: 14.19 -> 14.37 | Hurt |
| repeat D4-selected data for 2 epochs instead of single epoch on random new data | OPT-1.3B · Non-web snapshot perplexity | perplexity: 16.27 -> 16.10 | Improved |
| apply data selection (SemDeDup, SSL Prototypes or D4) | OPT-1.3B · web snapshot validation sets (C4, CC-dedup, CommonCrawl) | perplexity | Hurt |
| add re-clustering after SemDeDup before SSL Prototypes (ablation) | D4 · Instruction + Answers and Non-web perplexity | perplexity | Improved |
| combine SemDeDup with SSL Prototypes (D4) instead of either alone (ablation) | data selection · Instruct OPT validation perplexity, 1.3B | perplexity | Improved |
| vary number of K-Means clusters (ablation) | SSL Prototypes · 125M OPT on C4 | perplexity change | No clear effect |
| increase model scale from 125M to 6.7B | D4 · efficiency gain | efficiency gain: est. 20% for LLama-65B, 22% for OPT-175B | Improved |
Closest papers by meaning
- DataMan: Data Manager for Pre-training Large Language Models (ICLR 2025 · similarity 0.91)
- Improving Pretraining Data Using Perplexity Correlations (ICLR 2025 · similarity 0.91)
- Get more for less: Principled Data Selection for Warming Up Fine-Tuning in LLMs (ICLR 2024 · similarity 0.91)
- Fixing It in Post: A Comparative Study of LLM Post-Training Data Quality and Model Performance (NeurIPS 2025 · similarity 0.91)
- Dataset Decomposition: Faster LLM Training with Variable Sequence Length Curriculum (NeurIPS 2024 · similarity 0.91)
- Rethinking Data Curation in LLM Training: Online Reweighting Offers Better Generalization than Offline Methods (ICLR 2026 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
