NTQS Leaderboard · 纳西语翻译模型排行榜

The NTQS (Naxi Translation Quality Score) ranks AI language models on their ability to translate Naxi — a critically under-resourced Tibeto-Burman language. Scores combine tonal accuracy, semantic fidelity, and output completeness into a single 0–100 metric.

NTQS排行榜评估AI模型翻译纳西语的能力，综合声调准确率、语义忠实度和输出完整度三项指标，给出0–100的评分。

How NTQS Works · 评分机制

1. Tone accuracy (25%) — does the Pinyin output mark tones correctly? Naxi is tonal; a wrong tone changes the word.
2. Semantic fidelity (50%) — does the translation preserve meaning? Evaluated via native-speaker judgments.
3. Completeness (25%) — is the output a full, well-formed sentence?

Current Top Score · 当前最高分

82.3 NTQS — our best fine-tuned LoRA model, built on Gemma 2 9B. This represents a +41-point gain over the zero-shot frontier baseline (≈41.3 NTQS).

最高分82.3，为基于Gemma 2 9B的LoRA微调模型，比零样本基线提升41分。

Models Evaluated · 已评估模型

The leaderboard includes both zero-shot frontier models (GPT-4o, Claude, Gemini, Llama) and our domain-adapted LoRA fine-tuned models. See the models page for architecture and training details.