LoRA-adapted language models fine-tuned specifically for Naxi↔Chinese↔English translation. Trained on 60,000+ curated Naxi sentence pairs, published with full provenance — model weights, corpus, and evaluation methods are open.
Our best fine-tuned model scores 82.3 on the NTQS benchmark, a +41-point improvement over zero-shot frontier models. NTQS evaluates tonal accuracy (25%), semantic fidelity (50%), and completeness (25%).
Base: Qwen3-14B (earlier work used Gemma 2 9B) · Training: LoRA fine-tuning on 60,000+ parallel examples · Published on Hugging Face with full training provenance.
底座：Qwen3-14B · 方法：LoRA微调 · 已在Hugging Face公开发布。
NTQS Evaluation Criteria · 评估标准
Tone (25%) — correct Naxi tone marking (high/low/rising/mid)
Semantic (50%) — meaning preservation in translation
Completeness (25%) — full sentence output with no truncation
Weights, benchmarks and evaluation harnesses are published openly under
huggingface.co/Apeters247,
each with its own card, licence and provenance. Download counts below were current on
13 November 2026.
Models
naxi-qwen3-14b-v5
— LoRA adapter on Qwen3-14B. The model behind the 82.3 NTQS result;
234 downloads.
naxi-qwen3.5-4b-v6
— full fine-tune on Qwen3.5-4B-Base, Naxi/Chinese/English; small enough to run on
a laptop. 995 downloads.
naxi-qwen3.5-4b-GGUF
— the same model quantised to Q5_K_M for llama.cpp / Ollama.
109 downloads.
Datasets and benchmarks
naxi-mt-benchmark
— evaluation data and harness for Naxi ↔ Chinese ↔ English translation: 30 systems
on one frozen instrument with two complementary metrics. 96 downloads.
bodbench
— a balanced Tibetan→English benchmark: 400 held-out sentences, 200 modern and
200 classical, because registers are routinely mixed in reported results.
237 downloads.
tibetan-translation-benchmark
— Tibetan→English evaluation data with a leak-safe split, sized for comparing and
fine-tuning translation models. 245 downloads.
The Tibetan benchmarks are the instruments behind the
Anuvāda; the Naxi benchmark is the instrument
behind the NTQS leaderboard.