Research map / Large language models / Retrieval and knowledge
Cram Less to Fit More: Training Data Pruning Improves Memorization of Facts
Jiayuan Ye, Vitaly Feldman, Kunal Talwar
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 |
|---|---|---|---|
| add loss-based head selection (LossH-Wiki, alpha 0.2) | 110m model full-data training · Wikipedia test facts (110m) | fact accuracy (accurate fact count): 696.11 -> 929.25 | Improved |
| prune facts with loss-based selection instead of random pruning (alpha 0.2) (ablation) | random fact pruning · Wikipedia test facts (110m) | fact accuracy (accurate fact count): 647.33 -> 929.25 | Improved |
| prune facts with loss-based selection instead of random pruning (alpha 0.9) (ablation) | random fact pruning · Wikipedia test facts (110m) | fact accuracy (accurate fact count): 690.86 -> 733.23 | Improved |
| random fact pruning instead of full data (ablation) | 110m model full-data training · Wikipedia test facts (110m) | fact accuracy (accurate fact count): 696.11 -> 690.86 | No clear effect |
| loss-based selection at small alpha 0.1 vs full data | 110m model · Wikipedia test facts (110m) | fact accuracy (accurate fact count): 696.11 -> 835.81 | Improved |
| loss-based selection on NLU accuracy at alpha 0.2 | 110m model · Wikipedia NLU | NLU accuracy: 0.3636 -> 0.3604 | No clear effect |
| prune training data to alpha 0.2 vs training a 10x larger model on full data | 110m model · Wikipedia test facts (110m) | fact accuracy | Improved |
| prune data (alpha 0.5) on Knowledge-MMLU | 110m model · Knowledge-MMLU | accuracy | Mixed |
| use fact-level selection instead of token-level selection (ablation) | LossH/LossHF · synthetic phonebook | fact accuracy | Improved |
| use sum of loss instead of average loss over fact tokens (ablation) | LossH/LossHF · synthetic phonebook | rank correlation to fact weight-to-bits ratio | Improved |
| use online percentile threshold per batch (ablation) | LossH/LossHF · synthetic phonebook | correlation between loss and fact frequency | Improved |
| add flattening to head selection (LossHF) (ablation) | LossH · power-law phonebook | convergence speed | Mixed |
Closest papers by meaning
- How Do Large Language Models Acquire Factual Knowledge During Pretraining? (NeurIPS 2024 · similarity 0.88)
- Generalization v.s. Memorization: Tracing Language Models’ Capabilities Back to Pretraining Data (ICLR 2025 · similarity 0.87)
- Scalable Influence and Fact Tracing for Large Language Model Pretraining (ICLR 2025 · similarity 0.87)
- Understanding Finetuning for Factual Knowledge Extraction (ICML 2024 · similarity 0.87)
- Long-form factuality in large language models (NeurIPS 2024 · similarity 0.86)
- Mitigating Memorization in Language Models (ICLR 2025 · similarity 0.86)
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.
