CantoneseLLMChat-v1.0-72B
Cantonese LLM Chat v1.0 is the first generation Cantonese LLM from hon9kon9ize. Building upon the sucess of v0.5 preview, the model excels in Hong Kong related specific knowledge and Cantonese conversation.
Model description
Base model obtained via Continuous Pre-Training of Qwen 2.5 72B with 600 millions publicaly available Hong Kong news articles and Cantonese websites. Instructions fine-tuned model trained with a dataset consists of 75,000 instrutions pairs. 45,000 pairs were Cantonese insturctions generated by other LLMs and reviewed by humans.
The model trained with 16 Nvidia H100 96GB HBM2e GPUs on Genkai Supercomputer.
Basic Usage
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
model_id = "hon9kon9ize/CantoneseLLMChat-v1.0-72B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype=torch.bfloat16,
device_map="auto",
)
def chat(messages, temperature=0.9, max_new_tokens=200):
input_ids = tokenizer.apply_chat_template(conversation=messages, tokenize=True, add_generation_prompt=True, return_tensors='pt').to('cuda:0')
output_ids = model.generate(input_ids, max_new_tokens=max_new_tokens, temperature=temperature)
response = tokenizer.decode(output_ids[0][input_ids.shape[1]:], skip_special_tokens=False)
return response
prompt = "邊個係香港特首？"
messages = [
{"role": "system", "content": "you are a helpful assistant."},
{"role": "user", "content": prompt}
]
print(chat(messages)) # 香港特別行政區行政長官係李家超。<|im_end|>
Performance
Best in class open source LLM in understanding Cantonese and Hong Kong culture in the HK-Eval Benchmark. However, as one could observe, reasoning models have performed dramatically better than their counterparts. We are currently working on reasoning models for v2.
| Model | HK Culture (zero-shot) | Cantonese Linguistics |
|---|---|---|
| CantonesellmChat v0.5 6B | 52.0% | 12.8% |
| CantonesellmChat v0.5 34B | 72.5% | 54.5% |
| CantonesellmChat v1.0 3B | 56.0% | 45.7% |
| CantonesellmChat v1.0 7B | 60.3% | 46.5% |
| CantonesellmChat v1.0 32B | 69.8% | 52.7% |
| CantonesellmChat v1.0 72B | 75.4% | 59.6% |
| Llama 3.1 8B Instruct | 45.6% | 35.1% |
| Llama 3.1 70B Instruct | 63.0% | 50.3% |
| Qwen2.5 7B Instruct | 51.2% | 30.3% |
| Qwen2.5 32B Instruct | 59.9% | 45.1% |
| Qwen2.5 72B Instruct | 65.9% | 45.9% |
| Claude 3.5 Sonnet | 71.7% | 63.2% |
| DeepSeek R1 | 88.8% | 77.5% |
| Gemini 2.0 Flash | 80.2% | 75.3% |
| Gemini 2.5 Pro | 92.1% | 87.3% |
| GPT4o | 77.5% | 63.8% |
| GPT4o-mini | 55.6% | 57.3% |