Research map / Large language models / Language, safety and interpretability
Joint Selection for Large-Scale Pre-Training Data via Policy Gradient-based Mask Learning
Ziqing Fan, Yuqiao Xian, Yan Sun, Ke Shen, Li Shen
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 |
|---|---|---|---|
| select with joint quality+diversity mask learning (FineWeb-Mask) instead of raw FineWeb | 1.5B dense · 12-task average | average accuracy (%): 44.9 -> 48.1 | Improved |
| select with FineWeb-Mask instead of raw FineWeb | 7B MoE · 12-task average | average accuracy (%): 50.7 -> 52.6 | Improved |
| select with FineWeb-Mask instead of FineWeb-DCLM (best baseline) | 7B MoE · 12-task average | average accuracy (%): 52.2 -> 52.6 | Improved |
| select with FineWeb-Mask instead of FineWebPro (best baseline) | 1.5B dense · 12-task average | average accuracy (%): 47.2 -> 48.1 | Improved |
| diversity-only selection (Semdedup) instead of raw FineWeb | 1.5B dense · 12-task average | average accuracy (%): 44.9 -> 43.8 | Hurt |
| quality-only selection (FineWeb-Edu) instead of raw FineWeb | 1.5B dense · 12-task average | average accuracy (%): 44.9 -> 46.8 | Improved |
| add quality-aware pruning (ablation) | DiSF mask selection · 12-task average | average accuracy (%): 44.3 -> 45.0 | Improved |
| add quality-aware initialization on top of pruning (ablation) | DiSF mask selection · 12-task average | average accuracy (%): 45.0 -> 45.1 | Improved |
| add DiSF diversity to joint objective (ablation) | DataMask joint learning · 12-task average | average accuracy (%) | Hurt |
| add Pair-wise Similarity or Facility Location to quality score (ablation) | DataMask joint learning · 12-task average (1.5B, 400B tokens) | average accuracy (%) | Improved |
| use FineWeb-Mask instead of FineWeb-Edu for continual pre-training plus SFT | 7B MoE · 12-task average | average accuracy (%): 59.1 -> 60.2 | Improved |
| use too small group number G (ablation) | DataMask · optimized diversity value | optimized score | Hurt |
Closest papers by meaning
- Towards Scalable Data Diversification for Language Model Pretraining via Leverage Score Sampling (NeurIPS 2026 · similarity 0.88)
- Predictive Data Selection: The Data That Predicts Is the Data That Teaches (ICML 2025 · similarity 0.88)
- Improving Pretraining Data Using Perplexity Correlations (ICLR 2025 · similarity 0.88)
- Enhancing Large Language Model Performance with Gradient-Based Parameter Selection (AAAI 2025 · similarity 0.87)
- DsDm: Model-Aware Dataset Selection with Datamodels (ICML 2024 · similarity 0.87)
- Harnessing Diversity for Important Data Selection in Pretraining Large Language Models (ICLR 2025 · 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.
