Qwen3-1.7B-Base-Nuosu-LoRA

Qwen3-1.7B-Base Nuosu Translation LoRA

This repository contains a research-preview LoRA adapter for translation between Standard Liangshan Yi (Nuosu), Chinese, and English. It is based on Qwen/Qwen3-1.7B-Base and is not a standalone model.

本仓库发布的是面向凉山规范彝文（诺苏语）的研究预览版翻译 LoRA。它基于 Qwen/Qwen3-1.7B-Base，必须与基础模型一起加载，不是完整模型，也不是通用对话模型。

Intended use / 适用范围

• Chinese-to-Nuosu, Nuosu-to-Chinese, and English-related translation research.
• Reproducible evaluation and continued corpus/model development.
• Explicit translation instructions with deterministic decoding.

该模型适合明确指定翻译方向的研究和实验。它没有经过足够的真实多轮对话训练，不适合作为 通用聊天助手，也不应被用于需要专业母语翻译质量的场景。

Training data / 训练数据

The adapter was trained through OCR-ground-truth continued pretraining, dictionary SFT, and short/long translation SFT. Sources used by this run:

• NuosuBench
• NuosuBburma OCR Evaluation Set
• Yi–Chinese Electronic Dictionary

NuosuBench was split into fixed train, validation, and held-out research-test partitions. The held-out partition was not used for gradient training.

Evaluation / 评测

Deterministic greedy decoding was used with max_new_tokens=96.

The 192-row validation gate had no empty outputs or replacement characters and a 1.04% length-truncation rate. On the full research test, the empty-output rate was 0.009%, the replacement-character rate was 1.77%, and the length-truncation rate was 3.72%.

Automatic scores do not establish linguistic correctness. Native-speaker review remains necessary.

Known limitations / 已知问题

• Some difficult Chinese-to-Nuosu prompts trigger repetitive Yi syllable loops until the generation limit.
• A small number of truncated generations end with the Unicode replacement character (�).
• Rare expressions, culturally specific phrases, and long translations remain unreliable.
• The training distribution is translation-heavy. Open-ended writing and multi-turn dialogue behavior are not reliable.
• The model has not received comprehensive native-speaker evaluation.

本模型最明显的问题是少量困难汉译彝输入会产生重复循环，直到达到生成长度上限。因此本版 仅作为 research preview 发布，不应描述为成熟的彝语翻译系统。

Usage / 使用方法

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

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

tokenizer = AutoTokenizer.from_pretrained(adapter_id)
base = AutoModelForCausalLM.from_pretrained(
    base_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(base, adapter_id)
model.eval()

messages = [
    {
        "role": "user",
        "content": "请把“欢迎来到凉山”翻译成凉山规范彝文，只输出译文。",
    }
]
prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
stop_ids = sorted(
    {
        tokenizer.eos_token_id,
        tokenizer.convert_tokens_to_ids("<|im_end|>"),
        tokenizer.convert_tokens_to_ids("<|endoftext|>"),
    }
)

with torch.inference_mode():
    output = model.generate(
        **inputs,
        max_new_tokens=96,
        do_sample=False,
        eos_token_id=stop_ids,
        pad_token_id=tokenizer.pad_token_id,
    )

print(
    tokenizer.decode(
        output[0, inputs["input_ids"].shape[-1]:],
        skip_special_tokens=True,
    )
)

Use explicit translation instructions and deterministic decoding. Stop and discard an output if it enters a repetition loop.

Training summary

• Base model: Qwen/Qwen3-1.7B-Base
• Adapter: LoRA, rank 32, alpha 64, dropout 0.05
• Precision: BF16
• Training topology: three GPUs with distributed data parallelism
• Stages: OCR CPT → dictionary SFT → short translation SFT → long translation SFT

Disclaimer / 免责声明

This artifact is provided for academic and research use. Users are responsible for validating generated translations and reviewing the terms of each cited data source. Do not use generated text as authoritative linguistic, educational, legal, medical, or public-service translation without qualified human review.