SPICE: Submodular Penalized Information-Conflict Selection for Efficient Large Language Model Training
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , , , , , , , , , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2026
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866911575606034432 |
|---|---|
| author | Chang, Powei Zhang, Jinpeng Chen, Bowen Wang, Chenyu Guo, Chenlu Zhang, Yixing Gao, Yukang Xiang, JianXiang Gao, Yue Sun, Chaoqun Chen, Yiyi Kong, Dongying |
| author_facet | Chang, Powei Zhang, Jinpeng Chen, Bowen Wang, Chenyu Guo, Chenlu Zhang, Yixing Gao, Yukang Xiang, JianXiang Gao, Yue Sun, Chaoqun Chen, Yiyi Kong, Dongying |
| contents | Information-based data selection for instruction tuning is compelling: maximizing the log-determinant of the Fisher information yields a monotone submodular objective, enabling greedy algorithms to achieve a $(1-1/e)$ approximation under a cardinality budget. In practice, however, we identify alleviating gradient conflicts, misalignment between per-sample gradients, is a key factor that slows down the decay of marginal log-determinant information gains, thereby preventing significant loss of information. We formalize this via an $\varepsilon$-decomposition that quantifies the deviation from ideal submodularity as a function of conflict statistics, yielding data-dependent approximation factors that tighten as conflicts diminish. Guided by this analysis, we propose SPICE, a conflict-aware selector that maximizes information while penalizing misalignment, and that supports early stopping and proxy models for efficiency. Empirically, SPICE selects subsets with higher log-determinant information than original criteria, and these informational gains translate into performance improvements: across 8 benchmarks with LLaMA2-7B and Qwen2-7B, SPICE uses only 10% of the data, yet matches or exceeds 6 methods including full-data tuning. This achieves performance improvements with substantially lower training cost. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2601_23155 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | SPICE: Submodular Penalized Information-Conflict Selection for Efficient Large Language Model Training Chang, Powei Zhang, Jinpeng Chen, Bowen Wang, Chenyu Guo, Chenlu Zhang, Yixing Gao, Yukang Xiang, JianXiang Gao, Yue Sun, Chaoqun Chen, Yiyi Kong, Dongying Machine Learning Artificial Intelligence Information-based data selection for instruction tuning is compelling: maximizing the log-determinant of the Fisher information yields a monotone submodular objective, enabling greedy algorithms to achieve a $(1-1/e)$ approximation under a cardinality budget. In practice, however, we identify alleviating gradient conflicts, misalignment between per-sample gradients, is a key factor that slows down the decay of marginal log-determinant information gains, thereby preventing significant loss of information. We formalize this via an $\varepsilon$-decomposition that quantifies the deviation from ideal submodularity as a function of conflict statistics, yielding data-dependent approximation factors that tighten as conflicts diminish. Guided by this analysis, we propose SPICE, a conflict-aware selector that maximizes information while penalizing misalignment, and that supports early stopping and proxy models for efficiency. Empirically, SPICE selects subsets with higher log-determinant information than original criteria, and these informational gains translate into performance improvements: across 8 benchmarks with LLaMA2-7B and Qwen2-7B, SPICE uses only 10% of the data, yet matches or exceeds 6 methods including full-data tuning. This achieves performance improvements with substantially lower training cost. |
| title | SPICE: Submodular Penalized Information-Conflict Selection for Efficient Large Language Model Training |
| topic | Machine Learning Artificial Intelligence |
| url | https://arxiv.org/abs/2601.23155 |