arXiv cs.CL· Vihindi Kotalawala, Nevidu Jayatilleke·· 9 天前AI 评分25
PlurVA-LLM 2026 Track-1：通过多语言微调与阈值校准实现 LLM 多元价值对齐
PlurVA-LLM-2026 Shared Task Track-1: Pluralistic Value Alignment in LLMs via Multilingual Fine-Tuning and Threshold Calibration
AI 导读
针对 PlurVA-LLM 2026 共享任务 Track-1 的中国、印尼、斯里兰卡多元价值对齐，该系统用 4-bit QLoRA 微调 Llama 3.1 8B Instruct，并结合选项置换增强、标注者投票扩展与 SinhalaMMLU 增强及条件阈值校准。最终准确率为中国 0.785、印尼 0.715、斯里兰卡 0.916，整体宏平均 0.805。
正文 · 原文
Abstract:We present our system for the PlurVA-LLM 2026 Shared Task Track-1, which focuses on pluralistic value alignment in the contexts of China, Indonesia, and Sri Lanka. For this resource-constrained track, we fine-tuned Llama 3.1 8B Instruct using 4-bit QLoRA. Our approach combines option-permutation augmentation for Chinese data, annotator vote expansion for Indonesian data, and binary reformulation with SinhalaMMLU augmentation for Sri Lankan data. We further applied conditional threshold calibration to the predictions for the Sri Lankan data. The final system achieved accuracies of 0.785 for Chinese, 0.715 for Indonesian, and 0.916 for Sri Lankan, resulting in an overall macro-average accuracy of 0.805.
| Comments: | 11 pages, 6 figures, 6 tables, Accepted paper at the first workshop on Pluralistic Value Alignment of LLMs @ AACL-IJCNLP 2026 |
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2609.32382 [cs.CL] |
| (or arXiv:2609.32382v2 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2609.32382 arXiv-issued DOI via DataCite |
Submission history
From: Nevidu Jayatilleke Mr. [view email]
[v1]
Sat, 26 Sep 2026 08:56:11 UTC (917 KB)
[v2]
Thu, 8 Oct 2026 11:56:29 UTC (950 KB)
来源：arXiv cs.CL · arxiv.org