Yvthyvq/liujgoj-cantonese-v1

Liujgoj Cantonese LLM (V1.0-Alpha)

🌟 Introduction

Liujgoj Cantonese LLM is a fine-tuned language model built for the Liujgoj (溜歌粵語) romanization system.

Liujgoj is more than a phonetic transcription method. It is an experimental writing system designed to represent spoken Cantonese through:

• word-based orthography
• Latin script writing
• efficient tone encoding
• native Cantonese expression

This model is trained to convert natural Cantonese Chinese text into standardized Liujgoj Romanization.

---

中文簡介

Liujgoj Cantonese LLM 係專為 溜歌粵語（Liujgoj）羅馬字系統 微調嘅語言模型。

Liujgoj 唔單止係拼音方案，而係一套以廣東話口語為核心、強調「單詞化」同「拉丁字母化」嘅書寫系統。

本模型主要用途係：

• 將廣東話漢字句子轉換為 Liujgoj 羅馬字
• 學習地道口語詞彙組合
• 處理溜歌獨特拼寫規則
• 推動廣東話 AI 書寫技術發展

---

🚀 Key Features

✅ Word-Based Orthography

Learns merged word forms instead of character-by-character output.

Examples:

• 食咗 → sikhzor
• 靚仔 → lengzair
• 做咩 → zouh mej

---

✅ Tone-as-Letter System

Supports Liujgoj tone letters:

• j
• r
• x
• q
• h

This allows compact tone representation without numbers.

---

✅ Native Cantonese Fluency

Training data is derived from authentic Hong Kong Cantonese dialogue, preserving:

• colloquial speech
• sentence particles
• slang usage
• real spoken rhythm

---

📊 Training Data

Dataset Size

30,082 high-quality instruction pairs

Sources

Curated from 60+ Hong Kong movie subtitle files (SRT) and converted through manual / semi-automatic annotation.

Format Example

{
  "instruction": "Convert Cantonese Chinese into Liujgoj Romanization",
  "input": "你食咗飯未啊？",
  "output": "Neiq sikhzor faanh meih aa?"
}

---

🧠 Base Model

• unsloth/Qwen2.5-7B-bnb-4bit

Fine-tuning Method

• LoRA
• Supervised Fine-Tuning (SFT)
• Unsloth optimized training pipeline

---

🛠️ Usage

Python Example

from unsloth import FastLanguageModel
import torch

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "yvthyvq/liujgoj-cantonese-lora",
    max_seq_length = 2048,
    load_in_4bit = True,
)

FastLanguageModel.for_inference(model)

prompt = """Below is an instruction that describes a task.

### Instruction:
Convert Cantonese Chinese into Liujgoj Romanization.

### Input:
{}

### Response:
"""

inputs = tokenizer(
    [prompt.format("你食咗飯未啊？")],
    return_tensors="pt"
).to("cuda")

outputs = model.generate(**inputs, max_new_tokens=64)

print(tokenizer.batch_decode(outputs))

---

💬 Example

Input

你食咗飯未啊？

Output

Neiq sikhzor faanh meih aa?

---

⚠️ Limitations

This is an Alpha release.

Current limitations may include:

• unseen slang words
• long-context instability
• occasional spelling inconsistency
• hallucination on unrelated tasks

Recommended primarily for:

• Cantonese romanization
• Liujgoj experiments
• linguistic research
• niche Cantonese NLP tasks

---

🗺️ Roadmap

V1.0 Alpha

• Initial public release
• Cantonese Hanzi → Liujgoj conversion

V2.0 Planned

• improved accuracy
• better segmentation
• stronger instruction following
• broader vocabulary coverage

Future Goals

• Liujgoj chat assistant
• speech alignment
• grammar tools
• full Cantonese-native LLM ecosystem

---

🏷️ Tags

cantonese yue romanization liujgoj lora qwen unsloth linguistics

---

🙌 Author

Created by Yvthyvq

Building language technology for Cantonese and Liujgoj.

---