Research map / Large language models / Language, safety and interpretability
Optimal Splitting of Language Models from Mixtures to Specialized Domains
Skyler Seto, Pierre Ablin, Anastasiia Filippova, Jiayuan Ye, Louis Béthune, Angelos Katharopoulos, David Grangier
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 |
|---|---|---|---|
| split into per-cluster models after pretraining on mixture, at optimal split point | 1.3B LM · Pile specialized domains | perplexity: 9.33% average relative improvement | Improved |
| use split model training at scaling-law optimal ts | 1.3B and 2.7B LMs · zero-shot QA (8 benchmarks) | accuracy: +1.5 to 2% | Improved |
| 2.7B split model vs pretrained 6.7B model | 2.7B split · zero-shot QA | accuracy: +0.6% | Improved |
| 1.3B split model vs pretrained 6.7B model | 1.3B split · zero-shot QA | accuracy | Hurt |
| continued pretraining from scratch with no shared pretraining (ablation) | split training · 1.3B at 120/240/360B tokens | loss and QA accuracy | Hurt |
| split too early or too late rather than at optimal point (ablation) | split training · phonebook memorization (0% and 50% overlap) | average fact accuracy: peak at 20%-40% pretraining share | Mixed |
| greedy splitting at Dsplit instead of horizon-aware optimal ts (ablation) | split training · scaling law prediction | optimal allocation | Hurt |
| route to only two domain models (task subset) instead of all models (ablation) | split model · zero-shot QA | accuracy: about 1% lower | Hurt |
| MMLU subset routing instead of full routing (ablation) | split model · zero-shot QA | accuracy: up to 3% lower | Hurt |
| use GTE-small instead of SentenceBert embeddings for clustering/routing (ablation) | split model · zero-shot QA | accuracy: 62.6 -> 62.76 | No clear effect |
| random routing instead of cluster routing (ablation) | split model · zero-shot QA | accuracy | Hurt |
| loss-based routing instead of SBert cluster routing (ablation) | split model · zero-shot QA | accuracy | No clear effect |
Closest papers by meaning
- Nexus: Same Pretraining Loss, Better Downstream Generalization via Common Minima (NeurIPS 2026 · similarity 0.90)
- Model Merging in Pre-training of Large Language Models (NeurIPS 2025 · similarity 0.89)
- Chameleon: A Flexible Data-mixing Framework for Language Model Pretraining and Finetuning (ICML 2025 · similarity 0.89)
- The Journey Matters: Average Parameter Count over Pre-training Unifies Sparse and Dense Scaling Laws (ICLR 2025 · similarity 0.89)
- Beyond Next Token Prediction: Patch-Level Training for Large Language Models (ICLR 2025 · similarity 0.89)
- TANDEM: Bi-Level Data Mixture Optimization with Twin Networks (NeurIPS 2025 · similarity 0.89)
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.
