Research map / Architectures and efficiency / RNNs and normalization
Scaling Data-Constrained Language Models
Niklas Muennighoff, Alexander M Rush, Boaz Barak, Teven Le Scao, Nouamane Tazi, Aleksandra Piktus, Sampo Pyysalo, Thomas Wolf, Colin Raffel
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 up to 4 epochs on repeated data instead of unique data | 8.7B GPT-2 style LM · C4 validation loss | validation loss: 0.5% higher validation loss | No clear effect |
| allocate compute by data-constrained law (fewer params, more epochs) vs Chinchilla allocation | 9.3e21 FLOPs budget, 25B unique tokens · C4 test loss | loss: 27% fewer parameters, better loss | Improved |
| fill up to 50% of data budget with Python code | 4.2B LM, 84B tokens · 19 NL tasks average | downstream score | No clear effect |
| fill more than 50% of data with code | 4.2B LM, 84B tokens · 19 NL tasks average | downstream score | Hurt |
| perplexity filtering (top 25% lowest perplexity) | 4.2B LM, 84B tokens · 19 NL tasks average | downstream score | Improved |
| deduplication filtering | 4.2B LM, 84B tokens · 19 NL tasks average | downstream score | No clear effect |
| repeat data beyond about 4 epochs | 4.2B LM · downstream average | downstream score | Hurt |
Closest papers by meaning
- Language models scale reliably with over-training and on downstream tasks (ICLR 2025 · similarity 0.91)
- Resolving Discrepancies in Compute-Optimal Scaling of Language Models (NeurIPS 2024 · similarity 0.91)
- Data Engineering for Scaling Language Models to 128K Context (ICML 2024 · similarity 0.91)
- Bridging Compute- and Data-Optimal Pretraining (NeurIPS 2026 · similarity 0.91)
- Data-Constrained Language Model Pretraining: Improved Regularization and Scaling Laws (NeurIPS 2026 · similarity 0.91)
- Position: Will we run out of data? Limits of LLM scaling based on human-generated data (ICML 2024 · 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.
