metadata

license: bigscience-bloom-rail-1.0
language:
  - zh

体验链接

🔗http://101.68.79.42:7861/

🚀更新

模型链接	训练的数据量	模型版本	备注
https://huggingface.co/yuanzhoulvpi/chinese_bloom_7b_chat	15w中文指令数据	v1
https://huggingface.co/yuanzhoulvpi/chinese_bloom_7b_chat_v2	150w条中文指令数据	v2	目前已经测试过效果，相较于v1，效果有所提升
https://huggingface.co/yuanzhoulvpi/chinese_bloom_7b_chat_v3	420w条中文指令数据	v3	目前效果还没测试，欢迎大家测试

介绍

✅ 对bloom-7b模型做了sft，本次版本为V2版本(使用了150w条有监督数据做sft)，相较于V1版本，效果更好！！！
🚀 训练代码和推理代码全部分享，可以查看链接https://github.com/yuanzhoulvpi2017/zero_nlp/tree/main/chinese_bloom

如何使用

from transformers import AutoModelForCausalLM, AutoTokenizer


checkpoint = "yuanzhoulvpi/chinese_bloom_7b_chat_v2"#"bigscience/bloomz-3b" #"bigscience/bloom-7b1"#  "output_dir/checkpoint-8260"#
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModelForCausalLM.from_pretrained(checkpoint).half().cuda()

PROMPT_DICT = {
    "prompt_input": (
        "Below is an instruction that describes a task, paired with an input that provides further context. "
        "Write a response that appropriately completes the request.\n\n"
        "### Instruction:\n{instruction}\n\n### Input:\n{input}\n\n### Response:"
    ),
    "prompt_no_input": (
        "Below is an instruction that describes a task. "
        "Write a response that appropriately completes the request.\n\n"
        "### Instruction:\n{instruction}\n\n### Response:"
    ),
}

from typing import Optional
def generate_input(instruction:Optional[str]= None, input_str:Optional[str] = None) -> str:
    if input_str is None:
        return PROMPT_DICT['prompt_no_input'].format_map({'instruction':instruction})
    else:
        return PROMPT_DICT['prompt_input'].format_map({'instruction':instruction, 'input':input_str})


for i in range(5):
    print("*"*80)

    inputs = tokenizer.encode(generate_input(instruction="你是谁"), return_tensors="pt")
    outputs = model.generate(inputs,num_beams=3,
                            max_new_tokens=512,
                            do_sample=False, 
                            top_k=10,
                            penalty_alpha=0.6,
                            temperature=0.8,
                            repetition_penalty=1.2)
    print(tokenizer.decode(outputs[0]))