Research map / Large language models / Language, safety and interpretability
GneissWeb: Preparing High Quality Data for LLMs at Scale
Hajar Emami Gohari, Swanand Ravindra Kadhe, Yousaf Shah, Constantin M Adam, Abdulhamid Adebayo, Praneet Adusumilli, Farhan Ahmed, Nathalie Baracaldo, Santosh Subhashrao Borse, Yuan-Chi Chang, Xuan-Hong Dang, Nirmit Desai and 20 more
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 GneissWeb instead of FineWeb-V1.1.0 | 1.4B model, 350B tokens · High-Signal avg | accuracy: 56.26 -> 58.40 | Improved |
| train on GneissWeb instead of FineWeb-V1.1.0 | 7B model, 350B tokens · High-Signal avg | accuracy: 64.61 -> 67.34 | Improved |
| train on GneissWeb instead of FineWeb-V1.1.0 | 3B model, 350B tokens · Extended avg | accuracy: 50.15 -> 52.1 | Improved |
| train on GneissWeb instead of DCLM-Baseline | 1.4B model, 350B tokens · High-Signal avg | accuracy: 61.36 -> 58.40 | Hurt |
| train on GneissWeb instead of FineWeb-Edu-Score-2 | 1.4B model, 350B tokens · World Knowledge | accuracy: 42.24 -> 41.09 | Hurt |
| add exact substring deduplication to FineWeb (ablation) | 1.4B model · High-Signal avg, 350B tokens | accuracy: 55.99 -> 57.39 | Improved |
| add category-aware readability score filter (ablation) | FineWeb-V1.1.0 · High-Signal avg, 35B tokens | accuracy: 51.94 -> 53.20 | Improved |
| add category-aware extreme-tokenized filter (ablation) | FineWeb-V1.1.0 · High-Signal avg, 35B tokens | accuracy: 51.94 -> 52.78 | Improved |
| extend lenient thresholds to news, business, personal finance categories (ablation) | readability/extreme-token filters · High-Signal avg | accuracy | Hurt |
| use GneissWeb ensemble rule instead of rule 4 (AND of fastText) (ablation) | ensemble quality filter · High-Signal avg, 1.4B 35B | accuracy: 51.29 -> 51.66 | Improved |
| use GneissWeb ensemble rule instead of rule 1 (ablation) | ensemble quality filter · High-Signal avg, 7B 100B | accuracy: 62.65 -> 63.09 | Improved |
| use ensemble rule 3 instead of FineWeb baseline (ablation) | ensemble quality filter · High-Signal avg, 1.4B 35B | accuracy: 50.74 -> 51.06 | No clear effect |
Closest papers by meaning
- Fixing It in Post: A Comparative Study of LLM Post-Training Data Quality and Model Performance (NeurIPS 2025 · similarity 0.91)
- LLMEval: A Preliminary Study on How to Evaluate Large Language Models (AAAI 2024 · similarity 0.89)
- The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale (NeurIPS 2024 · similarity 0.89)
- Enhancing Multilingual LLM Pretraining with Model-Based Data Selection (NeurIPS 2025 · similarity 0.89)
- tinyBenchmarks: evaluating LLMs with fewer examples (ICML 2024 · similarity 0.88)
- DataMan: Data Manager for Pre-training Large Language Models (ICLR 2025 · similarity 0.88)
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.
