CantoneseLLM-6B-preview202402/README.md

language, license, license_name, license_link, pipeline_tag, model-index

language	license	license_name	license_link	pipeline_tag	model-index
yue	other	yi-license	https://huggingface.co/01-ai/Yi-6B/blob/main/LICENSE	text-generation	name results CantoneseLLM-6B-preview202402 task dataset metrics source type name text-generation Text Generation name type config split args AI2 Reasoning Challenge (25-Shot) ai2_arc ARC-Challenge test num_few_shot 25 type value name acc_norm 55.63 normalized accuracy url name https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=hon9kon9ize/CantoneseLLM-6B-preview202402 Open LLM Leaderboard task dataset metrics source type name text-generation Text Generation name type split args HellaSwag (10-Shot) hellaswag validation num_few_shot 10 type value name acc_norm 75.8 normalized accuracy url name https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=hon9kon9ize/CantoneseLLM-6B-preview202402 Open LLM Leaderboard task dataset metrics source type name text-generation Text Generation name type config split args MMLU (5-Shot) cais/mmlu all test num_few_shot 5 type value name acc 63.07 accuracy url name https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=hon9kon9ize/CantoneseLLM-6B-preview202402 Open LLM Leaderboard task dataset metrics source type name text-generation Text Generation name type config split args TruthfulQA (0-shot) truthful_qa multiple_choice validation num_few_shot 0 type value mc2 42.26 url name https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=hon9kon9ize/CantoneseLLM-6B-preview202402 Open LLM Leaderboard task dataset metrics source type name text-generation Text Generation name type config split args Winogrande (5-shot) winogrande winogrande_xl validation num_few_shot 5 type value name acc 74.11 accuracy url name https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=hon9kon9ize/CantoneseLLM-6B-preview202402 Open LLM Leaderboard task dataset metrics source type name text-generation Text Generation name type config split args GSM8k (5-shot) gsm8k main test num_few_shot 5 type value name acc 30.71 accuracy url name https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=hon9kon9ize/CantoneseLLM-6B-preview202402 Open LLM Leaderboard

CantoneseLLM

This model is further pre-trained model based on 01-ai/Yi-6B with 800M tokens of Cantonese text compiled from various sources, including translated zh-yue Wikipedia, translated RTHK news datasets/jed351/rthk_news, Cantonese filtered CC100 and Cantonese textbooks generated by Gemini Pro.

This is a preview version, for experimental use only, we will use it to fine-tune on downstream tasks and evaluate the performance.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	56.93
ARC (25-shot)	55.63
HellaSwag (10-shot)	75.8
MMLU (5-shot)	63.07
TruthfulQA (0-shot)	42.26
Winogrande (5-shot)	74.11
GSM8K (5-shot)	30.71

Usage

from transformers import AutoTokenizer, AutoModelForMaskedLM

tokenizer = AutoTokenizer.from_pretrained("hon9kon9ize/CantoneseLLM-6B-preview202402")
model = AutoModelForMaskedLM.from_pretrained("hon9kon9ize/CantoneseLLM-6B-preview202402")

prompt = "歷經三年疫情，望穿秋水終於全面復常，隨住各項防疫措施陸續放寬以至取消，香港"

input_ids = tokenizer.encode(prompt, return_tensors="pt").to('cuda:0')
output = model.generate(input_ids, max_length=max_length, num_return_sequences=1, repetition_penalty=1.1, do_sample=True, temperature=temperature, top_k=50, top_p=0.95)
output = tokenizer.decode(output[0], skip_special_tokens=True)

# output: 歷經三年疫情，望穿秋水終於全面復常，隨住各項防疫措施陸續放寬以至取消，香港旅遊業可謂「起死回生」。
# 不過，旅遊業嘅復蘇之路並唔順利，香港遊客數量仍然遠低於疫前水平，而海外旅客亦只係恢復到疫情前約一半。有業界人士認為，當局需要進一步放寬入境檢疫措施，吸引更多國際旅客來港，令旅遊業得以真正復甦。

Limitation and Bias

The model is intended to use for Cantonese language understanding and generation tasks, it may not be suitable for other Chinese languages. The model is trained on a diverse range of Cantonese text, including news, Wikipedia, and textbooks, it may not be suitable for informal or dialectal Cantonese, it may contain bias and misinformation, please use it with caution.

We found the model is not well trained on the updated Hong Kong knowledge, it may due to the corpus is not large enough to brainwash the original model. We will continue to improve the model and corpus in the future.