Research map / Large language models / Instruction and fine-tuning
LangGPS: Language Separability Guided Data Pre-Selection for Joint Multilingual Instruction Tuning
Yangfan Ye, Xiaocheng Feng, Xiachong Feng, Lei Huang, Weitao Ma, Qichen Hong, Yunfei Lu, Duyu Tang, Dandan Tu, Bing Qin
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 |
|---|---|---|---|
| add LangGPS pre-selection to LESS | LESS (LLaMA-3.1-8B) · six benchmarks | relative change vs full data: +4.90% -> +6.37% | Improved |
| add LangGPS pre-selection to random selection | Rand (LLaMA-3.1-8B) · XStoryCloze | accuracy: 55.4 -> 64.6 | Improved |
| add LangGPS pre-selection to LESS | LESS (LLaMA-3.1-8B) · MKQA | accuracy: 17.8 -> 16.9 | Hurt |
| add LangGPS pre-selection to KMC | KMC (Qwen2.5-7B) · XStoryCloze | accuracy: 70.6 -> 73.7 | Improved |
| add LangGPS pre-selection to DSIR | DSIR (Qwen2.5-7B) · MMMLU | accuracy: 45.4 -> 48.7 | Improved |
| add LangGPS pre-selection to MTLD | MTLD (Qwen2.5-7B) · XLSum | ROUGE-L: 22.2 -> 21.8 | Hurt |
| add LangGPS pre-selection to Coh | Coh (LLaMA-3.1-8B) · XQuAD | accuracy: 63.2 vs 64.5 | Hurt |
| choose LESS over random selection | LESS (Qwen2.5-7B) · six benchmarks | relative change vs full data: -7.35% vs +4.18% | Hurt |
| use balanced separability curriculum instead of random shuffle (ablation) | full-data SFT (LLaMA-3.1-8B) · XStoryCloze | accuracy: 56.8 -> 68.8 | Improved |
| use balanced separability curriculum instead of random shuffle (ablation) | full-data SFT (Qwen2.5-7B) · XNLI | accuracy: 45.3 -> 50.8 | Improved |
| use ascending separability curriculum instead of random shuffle (ablation) | full-data SFT (Qwen2.5-7B) · XNLI | accuracy: 45.3 -> 36.4 | Hurt |
| vary pre-selection ratio rho from 10% to 100% (ablation) | LangGPS+Rand/LESS (LLaMA-3.1-8B) · MMMLU, XLSum (5%) | accuracy / ROUGE-L | Mixed |
Closest papers by meaning
- XIFBench: Evaluating Large Language Models on Multilingual Instruction Following (NeurIPS 2025 · similarity 0.90)
- MCIF: Multimodal Crosslingual Instruction-Following Benchmark from Scientific Talks (ICLR 2026 · similarity 0.88)
- Exploring Polyglot Harmony: On Multilingual Data Allocation for Large Language Models Pretraining (NeurIPS 2025 · similarity 0.88)
- XCOT: Cross-lingual Instruction Tuning for Cross-lingual Chain-of-Thought Reasoning (AAAI 2025 · similarity 0.87)
- MuRating: A High Quality Data Selecting Approach to Multilingual Large Language Model Pretraining (NeurIPS 2025 · similarity 0.87)
- Language Imbalance Driven Rewarding for Multilingual Self-improving (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.
