This model is based on bigscience/bloomz-7b1-mt. To make it more accessible and efficient for certain Chinese , we have pruned its original vocabulary from 250,880 tokens to 46,145 tokens using Chinese corpus data as follow bloom-6b4-zh. This reduction in vocabulary size has helped to significantly reduce the GPU memory usage required to run the model. As a result, the total number of parameters in the model is now 6 billion 4.

基于 bigscience/bloomz-7b1-mt，修建embeddings层到 46145，主要保留中文相关的tokens映射。修建后参数为6B4。

How to use

from transformers import BloomTokenizerFast, BloomForCausalLM

tokenizer = BloomTokenizerFast.from_pretrained('enze/bloomz-6b4-zh')
model = BloomForCausalLM.from_pretrained('enze/bloomz-6b4-zh')

print(tokenizer.batch_decode(model.generate(tokenizer.encode('中国的首都是', return_tensors='pt'))))