BELLE icon indicating copy to clipboard operation
BELLE copied to clipboard

BELLE: Be Everyone's Large Language model Engine(开源中文对话大模型)

Results 163 BELLE issues
Sort by recently updated
recently updated
newest added

I transformed belle-whisper-large-v2 by ctranslate2, the model size is almost same as faster-whisper-large-v2. But when the word_timestamp parameter is True, Belle took much more time(at least 3x, sometimes 10x) than...

Can we continually fine-tune on belle-whisper-large-v2-zh by own data?

请问我想生成“安全隐私”这方面的数据集,是只需要运行1.5M文件夹下的指令即可吗?关于prompt的编写有什么好的建议吗?我看你们在hugging face上也开源了“数学题”这类的数据集,我想知道是如何提示大模型来生成的?还有我想问一下运行该文件后会生成1.5M这么大的数据吗?大约需要多大内存和显存,我后续可能要租服务器来运行该项目,本机配置太低,因此先来问一下;感谢答复!

### Description Inference speed is too slow and positively related to max_new_tokens length. For example, I set max_new_tokens=1000, which would take almost 30s~40s with A100 ### Background After loading a...

tothemoon:latest这个镜像里存在一些问题: 1.flash-attn的版本应该不OK,下面的flash_attn_varlen_func找不到 try: from flash_attn.flash_attn_interface import flash_attn_varlen_func except ImportError: flash_attn_varlen_func = None print( "Warning: import flash_attn fail, please install FlashAttention " "https://github.com/Dao-AILab/flash-attention" ) 2. 升级flash_attn版本到最新后,出现so里面找不到符号问题

![image](https://github.com/LianjiaTech/BELLE/assets/19919798/b76214f1-6640-45de-abb4-5dbcb85edcb4)

看到数据处理函数aishell.py中似乎并没有对一个长句子进行短句时间戳分段,而是从零开始到音频结束作为一整个时间戳 ![data_process](https://github.com/LianjiaTech/BELLE/assets/145308216/b45404da-ed23-45e6-b6ba-6c6e0f60da7d) 所以这个Belle-whisper-large-v2-zh模型的分句效果不太好是因为数据没进行分句的原因吗

The prompt I am using now is: 人设: 景芳:24岁,高校图书管理员。宅女,不喜欢户外活动。 HUMAN:老师,你叫什么名字啊? 景芳-Assistant:我的名字叫景芳 HUMAN:哦哦我是新来的学生,以后希望能和老师好好相处。 景芳-Assistant: However, the model continues to output irrelevant English words. I've checked the whole pipeline, and it seems...