71 lines
2.4 KiB
Markdown
71 lines
2.4 KiB
Markdown
|
---
|
|||
|
|
license: apache-2.0
|
|||
|
|
base_model: Yvthyvq/cantonese-qwen3-8b-base
|
|||
|
|
tags:
|
|||
|
|
- cantonese
|
|||
|
|
- orthography
|
|||
|
|
- liujgoj
|
|||
|
|
- speech-first
|
|||
|
|
- sft
|
|||
|
|
- text-generation
|
|||
|
|
pipeline_tag: text-generation
|
|||
|
|
---
|
|||
|
|
# cantonese-qwen3-8b-instruct
|
|||
|
|
|
|||
|
|
融入純粹粵語語音 manifold（語言流形）嘅多任務 Instruct 大模型。本模型基於 cantonese-qwen3-8b-base 進行深度大規模指令微調（SFT），專注於拉丁化粵語書寫系統 —— **Liujgoj 溜歌粵語**。
|
|||
|
|
|
|||
|
|
透過將音素同語義對齊，本模型擺脫咗傳統漢字對粵語思維嘅束縛，為語音原生（Speech-native）AI 奠定強大嘅文本語意底座。
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## 🚀 模型亮點
|
|||
|
|
|
|||
|
|
* **純語音導向（Phonology-first）**：基於 Liujgoj 拉丁化正詞法（Tone-as-letter 字母表調法：j, r, x, q, h），繞過漢字表意限制，實現更高效嘅 AI 語意向量建模。
|
|||
|
|
* **強大數據錘煉**：精選超過多部經典香港電影對白、高頻粵語詞庫（約 13,000 詞），精心策劃超 14 萬行高質量多任務對話與指令對（Instruction pairs）進行全參數/大窗口微調。
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## 🛠️ 快速部署與使用
|
|||
|
|
|
|||
|
|
### 1. 使用 Hugging Face 官方最新 `hf` 工具下載
|
|||
|
|
|
|||
|
|
由於模型採用最新 high-performance 傳輸架構，推薦使用最新 `hf` 工具進行下載（速度極快）：
|
|||
|
|
|
|||
|
|
```bash
|
|||
|
|
export HF_XET_HIGH_PERFORMANCE=1
|
|||
|
|
hf download Yvthyvq/cantonese-qwen3-8b-instruct --local-dir ./cantonese-qwen3-8b-instruct
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## 🚀 快速開始 / Quick Start
|
|||
|
|
|
|||
|
|
### 1. 安裝與環境變量配置
|
|||
|
|
```bash
|
|||
|
|
export HF_XET_HIGH_PERFORMANCE=1
|
|||
|
|
export HF_ENDPOINT=[https://hf-mirror.com](https://hf-mirror.com)
|
|||
|
|
|
|||
|
|
### 2. Python 檔案推演 (使用 Transformers 載入)
|
|||
|
|
```python
|
|||
|
|
import torch
|
|||
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|||
|
|
|
|||
|
|
model_id = "Yvthyvq/cantonese-qwen3-8b-instruct"
|
|||
|
|
|
|||
|
|
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
|||
|
|
model = AutoModelForCausalLM.from_pretrained(
|
|||
|
|
model_id,
|
|||
|
|
torch_dtype=torch.bfloat16,
|
|||
|
|
device_map="auto"
|
|||
|
|
)
|
|||
|
|
|
|||
|
|
# 提示詞請盡量使用 Liujgoj 粵語拼寫或地道口語進行互動
|
|||
|
|
messages = [
|
|||
|
|
{"role": "user", "content": "Neiq hour, neiq horyiq zouh dij mej?"} # 範例
|
|||
|
|
]
|
|||
|
|
|
|||
|
|
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
|||
|
|
inputs = tokenizer([prompt], return_tensors="pt").to(model.device)
|
|||
|
|
|
|||
|
|
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.7)
|
|||
|
|
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
|