This repository contains a research LoRA adapter for Chinese–Standard Liangshan Yi (Nuosu) and related short-form translation experiments. It is an adapter-only release; obtain the base model separately.

Intended use

Chinese ↔ Standard Liangshan Yi translation research;
short dictionary and sentence translation experiments;
reproducible low-resource language adaptation studies.

This is not a production translation system. The full held-out evaluation is weak on the heterogeneous research test distribution, and the fixed gate was run under an explicitly recorded waiver. Native-speaker review is required before making semantic or orthographic claims.

本版本不宣称覆盖全部彝语方言，也不适合高风险或未经审核的正式翻译。

Base model and reproducibility

Base model: Qwen/Qwen3-1.7B
Base revision: 70d244cc86ccca08cf5af4e1e306ecf908b1ad5e
Training code revision: 90b7d6c3d71e025e1336a2a585389f1dedab9b6f
Chat entrypoint fix: 3c7f17f012e5a483c367ff7b5b16e905e1b2c7dd
Dataset projection: nuosu-mt-clean-recover-v20260808
Seed: 42
LoRA: rank 64, alpha 128, dropout 0.05, all-linear targets
Training: BF16, one SFT epoch, completion-only loss

The tokenizer adds 1,203 Yi syllable/radical tokens (vocabulary size 152,872); new token rows are initialized from the original subtoken embeddings and trained together with LoRA.

Data

The target-only MT projection contains:

Split	Records	Notes
train	159,083	94,532 lexicon, 16,077 published, 39,512 sentence, 8,962 short
validation	7,131	held-out validation projection
research test	8,558	full held-out generation test

Training data were cleaned by dropping exact meta-evaluation verdict targets and recovering usable corrected-translation prefixes. The release contains model files and evaluation metadata, not the source corpora. Source licensing, attribution, and redistribution conditions remain applicable.

Evaluation

Full held-out generation (8,558 records, greedy, no_think):

Metric	Overall	Yi-target
Compact exact match	3.76%	2.47%
chrF2	10.93	13.70
Reference contained	5.68%	3.35%
Replacement-character rate	0.11%	0.33%

The 256-record gate reached 57.42% overall exact match, 60.68 chrF2 and 48.44% Yi exact match, but did not satisfy the strict gate thresholds; the waiver is included under provenance/GATE_WAIVER.

Usage

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_id = "Qwen/Qwen3-1.7B"
adapter_id = "NiceAsiv/Qwen3-1.7B-Nuosu-MT"

tokenizer = AutoTokenizer.from_pretrained(adapter_id)
base = AutoModelForCausalLM.from_pretrained(
    base_id, torch_dtype="auto", device_map="auto"
)
base.resize_token_embeddings(len(tokenizer), pad_to_multiple_of=64)
model = PeftModel.from_pretrained(base, adapter_id)

messages = [{
    "role": "user",
    "content": "请将以下中文翻译为凉山规范彝文。只输出译文，不要解释。\n我今天去学校。",
}]
prompt = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True,
    enable_thinking=False,
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(
    **inputs, do_sample=False, max_new_tokens=256,
    eos_token_id=tokenizer.eos_token_id,
)
print(tokenizer.decode(output[0, inputs["input_ids"].shape[-1]:],
                       skip_special_tokens=True).strip())

The repository's chat command defaults to the same deterministic no_think mode. Use --thinking-mode thinking only when deliberately testing reasoning.

Citation

bibtex

@software{axi2026nuosumt,
  author      = {Wuhe Axi},
  title       = {Qwen3-1.7B Nuosu MT LoRA},
  year        = {2026},
  institution = {Xi'an Jiaotong University},
  url         = {https://huggingface.co/NiceAsiv/Qwen3-1.7B-Nuosu-MT}
}

Please also cite the training code and the separately maintained corpus: