Research map / Large language models / Language, safety and interpretability
The RefinedWeb Dataset for Falcon LLM: Outperforming Curated Corpora with Web Data Only
Guilherme Penedo, Quentin Malartic, Daniel Hesslow, Ruxandra Cojocaru, Hamza Alobeidli, Alessandro Cappelli, Baptiste Pannier, Ebtesam Almazrouei, Julien Launay
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 |
|---|---|---|---|
| Train on RefinedWeb alone instead of The Pile | 1B model at 27GT · small-agg | zero-shot accuracy: 53.4% -> 56.2% | Improved |
| Train on RefinedWeb alone instead of C4 | 3B model at 60GT · small-agg | zero-shot accuracy: 59.6% -> 59.8% | Improved |
| Add filtering to RefinedWeb raw data (ablation) | RefinedWeb, 1B@27GT · small-agg | zero-shot accuracy: 52.7% -> 54.3% | Improved |
| Add final deduplication after filtering (ablation) | RefinedWeb, 1B@27GT · small-agg | zero-shot accuracy: 54.3% -> 56.2% | Improved |
| Apply MDR filtering (ablation) | OSCAR-21.09 · small-agg | zero-shot accuracy: 55.0% -> 55.4% | Improved |
| Apply MDR filtering (ablation) | OSCAR-22.01 · small-agg | zero-shot accuracy: 52.7% -> 52.3% | Hurt |
| Apply MDR filtering (ablation) | C4 · small-agg | zero-shot accuracy: 55.7% -> 56.2% | Improved |
| Apply MDR filtering (ablation) | The Pile · small-agg | zero-shot accuracy: 53.4% -> 54.2% | Improved |
| Apply stringent MDR deduplication (ablation) | OSCAR-21.09, OSCAR-22.01, C4, and The Pile · small-agg | zero-shot accuracy: 55.0% -> 55.6%; 52.7% -> 55.6%; 55.7% -> 55.9%; 53.4% -> 54.5% | Improved |
| Combine MDR filtering and deduplication (ablation) | OSCAR-21.09 · small-agg | zero-shot accuracy: 55.0% -> 55.5% | Improved |
| Combine MDR filtering and deduplication (ablation) | OSCAR-22.01 · small-agg | zero-shot accuracy: 52.7% -> 55.4% | Improved |
| Combine MDR filtering and deduplication (ablation) | C4 · small-agg | zero-shot accuracy: 55.7% -> 56.4% | Improved |
Closest papers by meaning
- The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale (NeurIPS 2024 · similarity 0.91)
- Organize the Web: Constructing Domains Enhances Pre-Training Data Curation (ICML 2025 · similarity 0.89)
- RedPajama: an Open Dataset for Training Large Language Models (NeurIPS 2024 · similarity 0.89)
- SWEb: A Large Web Dataset for the Scandinavian Languages (ICLR 2025 · similarity 0.87)
- The Common Pile v0.1: An 8TB Dataset of Public Domain and Openly Licensed Text (NeurIPS 2025 · similarity 0.87)
- Towards Scalable Data Diversification for Language Model Pretraining via Leverage Score Sampling (NeurIPS 2026 · similarity 0.87)
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.
