Research map / Graph learning / Graph algorithms and clustering
D$^3$: Dynamic Directional Graph-Constrained Data Scheduling for LLM Training
Yuanjian Xu, Jianing Hao, Guang Zhang, Zhong Li
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 |
|---|---|---|---|
| replace Random-Swap Refinement with Row-Sum solver (ablation) | D3 · SlimPajama GPT-2 Medium | perplexity: 3.04 -> 2.97 | Improved |
| replace Random-Swap Refinement with Greedy Insertion solver (ablation) | D3 · SlimPajama GPT-2 Medium | perplexity: 3.17 -> 2.97 | Improved |
| order batches randomly instead of D3 scheduling | D3 · SlimPajama GPT-2 Medium GPU hours | GPU hours: 5.53 -> 10.45 | Hurt |
| use diagonal Hessian instead of Fisher approximation for influence (ablation) | D3 · SlimPajama GPT-2 Medium | perplexity | Improved |
| use diagonal Hessian influence estimation (vs baseline) (ablation) | D3 · SlimPajama GPT-2 Medium | perplexity: 5.1% reduction; 89.0% overhead | Improved |
| tune gradient projection dimension to k=3500 (ablation) | D3 · SlimPajama GPT-2 Medium | perplexity: 8.1% reduction | Improved |
| increase scheduling window to L=60 (ablation) | D3 · SlimPajama GPT-2 Medium | perplexity: 9.3% gain | Improved |
| use D3 instead of Dynamic Loss | D3 · SlimPajama PPL | perplexity: 3.10 -> 2.97 | Improved |
| use D3 instead of Dynamic Loss | D3 · PIQA | accuracy: 56.8 -> 55.7 | Hurt |
| use D3 instead of uniform sampling | D3 · OBQA | accuracy: 11.7 -> 15.2 | Improved |
| use D3 instead of fixed curriculum loss low->high | D3 · SFT Avg, Llama-3-8B | average score: 46.2 -> 49.6 | Improved |
| use D3 instead of loss high->low curriculum | D3 · SFT Avg, Llama-3-8B | average score: 34.9 -> 49.6 | Improved |
Closest papers by meaning
- GraphRouter: A Graph-based Router for LLM Selections (ICLR 2025 · similarity 0.88)
- BLADE: Scalable Bi-level Adaptive Data Selection for LLM Training (NeurIPS 2026 · similarity 0.87)
- Rethinking Data Curation in LLM Training: Online Reweighting Offers Better Generalization than Offline Methods (ICLR 2026 · similarity 0.87)
- LLM Data Selection and Utilization via Dynamic Bi-level Optimization (ICML 2025 · similarity 0.87)
- Unveiling the Secret Recipe: A Guide For Supervised Fine-Tuning Small LLMs (ICLR 2025 · similarity 0.87)
- Dynamic Loss-Based Sample Reweighting for Improved Large Language Model Pretraining (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.
