313 lines
14 KiB
Markdown
313 lines
14 KiB
Markdown
# Llama-3-8B-Instruct-Chinese-chat
|
||
|
||
|
||
Llama-3-8B-Instruct in Chinese 自己微调版本
|
||
|
||
### 训练可用数据整理
|
||
|
||
| 数据集 | 介绍 |
|
||
|----------------------------------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
|
||
| [firefly-train-1.1M](https://huggingface.co/datasets/YeungNLP/firefly-train-1.1M) | 包含了23种常见的中文NLP任务的数据,并且构造了许多与中华文化相关的数据,如对联、作诗、文言文翻译、散文、金庸小说等。对于每个任务,由人工书写若干种指令模板,保证数据的高质量与丰富度,数据量为115万。 |
|
||
| [moss-003-sft-data](https://huggingface.co/datasets/YeungNLP/moss-003-sft-data) | 由复旦大学MOSS团队开源的中英文多轮对话数据,包含100万+数本。 |
|
||
| [school_math_0.25M](https://huggingface.co/datasets/YeungNLP/school_math_0.25M) | 由BELLE项目组开源的数学运算指令数据,包含25万条数问。 |
|
||
| [ruozhiba](https://huggingface.co/datasets/LooksJuicy/ruozhiba) | 弱智吧数据问答,据说比较锻炼模型的心智能力。 |
|
||
欢迎补充,要求中文且一问一答形式,适合用于提升llama3任务能力的数据集
|
||
|
||
### [github地址](https://github.com/Rookie1019/Llama-3-8B-Instruct-Chinese.git)
|
||
|
||
### 推荐微调工具
|
||
在此感谢以下项目,提供了许多优秀的中文微调工具,供大家参考:
|
||
- Firefly - https://github.com/yangjianxin1/Firefly
|
||
- LLaMA-Factory - https://github.com/hiyouga/LLaMA-Factory.git
|
||
|
||
|
||
### Chat版模型下载
|
||
- Instruct + 继续中文sft版
|
||
- [huggingface地址](https://huggingface.co/Rookie/Llama-3-8B-Instruct-Chinese)
|
||
|
||
### 模型量化加速、部署
|
||
|
||
### 模型使用
|
||
|
||
默认情况下直接运行以下代码即可体验llama3中文对话,请自行修改`model_name_or_path`为你下载的模型路径
|
||
|
||
```python
|
||
from transformers import AutoTokenizer, AutoConfig, AddedToken, AutoModelForCausalLM, BitsAndBytesConfig
|
||
from peft import PeftModel
|
||
from dataclasses import dataclass
|
||
from typing import Dict
|
||
import torch
|
||
import copy
|
||
|
||
## 定义聊天模板
|
||
@dataclass
|
||
class Template:
|
||
template_name:str
|
||
system_format: str
|
||
user_format: str
|
||
assistant_format: str
|
||
system: str
|
||
stop_word: str
|
||
|
||
template_dict: Dict[str, Template] = dict()
|
||
|
||
def register_template(template_name, system_format, user_format, assistant_format, system, stop_word=None):
|
||
template_dict[template_name] = Template(
|
||
template_name=template_name,
|
||
system_format=system_format,
|
||
user_format=user_format,
|
||
assistant_format=assistant_format,
|
||
system=system,
|
||
stop_word=stop_word,
|
||
)
|
||
|
||
# 这里的系统提示词是训练时使用的,推理时可以自行尝试修改效果
|
||
register_template(
|
||
template_name='llama3',
|
||
system_format='<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\n{content}<|eot_id|>',
|
||
user_format='<|start_header_id|>user<|end_header_id|>\n\n{content}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n',
|
||
assistant_format='{content}<|eot_id|>',
|
||
system=None,
|
||
stop_word='<|eot_id|>'
|
||
)
|
||
|
||
|
||
## 加载模型
|
||
def load_model(model_name_or_path, load_in_4bit=False, adapter_name_or_path=None):
|
||
if load_in_4bit:
|
||
quantization_config = BitsAndBytesConfig(
|
||
load_in_4bit=True,
|
||
bnb_4bit_compute_dtype=torch.float16,
|
||
bnb_4bit_use_double_quant=True,
|
||
bnb_4bit_quant_type="nf4",
|
||
llm_int8_threshold=6.0,
|
||
llm_int8_has_fp16_weight=False,
|
||
)
|
||
else:
|
||
quantization_config = None
|
||
|
||
# 加载base model
|
||
model = AutoModelForCausalLM.from_pretrained(
|
||
model_name_or_path,
|
||
load_in_4bit=load_in_4bit,
|
||
trust_remote_code=True,
|
||
low_cpu_mem_usage=True,
|
||
torch_dtype=torch.float16,
|
||
device_map='auto',
|
||
quantization_config=quantization_config
|
||
)
|
||
|
||
# 加载adapter
|
||
if adapter_name_or_path is not None:
|
||
model = PeftModel.from_pretrained(model, adapter_name_or_path)
|
||
|
||
return model
|
||
|
||
## 加载tokenzier
|
||
def load_tokenizer(model_name_or_path):
|
||
tokenizer = AutoTokenizer.from_pretrained(
|
||
model_name_or_path,
|
||
trust_remote_code=True,
|
||
use_fast=False
|
||
)
|
||
|
||
if tokenizer.pad_token is None:
|
||
tokenizer.pad_token = tokenizer.eos_token
|
||
|
||
return tokenizer
|
||
|
||
## 构建prompt
|
||
def build_prompt(tokenizer, template, query, history, system=None):
|
||
template_name = template.template_name
|
||
system_format = template.system_format
|
||
user_format = template.user_format
|
||
assistant_format = template.assistant_format
|
||
system = system if system is not None else template.system
|
||
|
||
history.append({"role": 'user', 'message': query})
|
||
input_ids = []
|
||
|
||
# 添加系统信息
|
||
if system_format is not None:
|
||
if system is not None:
|
||
system_text = system_format.format(content=system)
|
||
input_ids = tokenizer.encode(system_text, add_special_tokens=False)
|
||
# 拼接历史对话
|
||
for item in history:
|
||
role, message = item['role'], item['message']
|
||
if role == 'user':
|
||
message = user_format.format(content=message, stop_token=tokenizer.eos_token)
|
||
else:
|
||
message = assistant_format.format(content=message, stop_token=tokenizer.eos_token)
|
||
tokens = tokenizer.encode(message, add_special_tokens=False)
|
||
input_ids += tokens
|
||
input_ids = torch.tensor([input_ids], dtype=torch.long)
|
||
|
||
return input_ids
|
||
|
||
|
||
def main():
|
||
model_name_or_path = 'NousResearch/Meta-Llama-3-8B'
|
||
template_name = 'llama3'
|
||
adapter_name_or_path = None
|
||
|
||
template = template_dict[template_name]
|
||
|
||
load_in_4bit = False
|
||
|
||
max_new_tokens = 500
|
||
top_p = 0.9
|
||
temperature = 0.35
|
||
repetition_penalty = 1.1
|
||
|
||
# 加载模型
|
||
print(f'Loading model from: {model_name_or_path}')
|
||
print(f'adapter_name_or_path: {adapter_name_or_path}')
|
||
model = load_model(
|
||
model_name_or_path,
|
||
load_in_4bit=load_in_4bit,
|
||
adapter_name_or_path=adapter_name_or_path
|
||
).eval()
|
||
tokenizer = load_tokenizer(model_name_or_path if adapter_name_or_path is None else adapter_name_or_path)
|
||
if template.stop_word is None:
|
||
template.stop_word = tokenizer.eos_token
|
||
stop_token_id = tokenizer.encode(template.stop_word, add_special_tokens=True)
|
||
assert len(stop_token_id) == 1
|
||
stop_token_id = stop_token_id[0]
|
||
|
||
history = []
|
||
|
||
query = input('# User:')
|
||
while True:
|
||
query = query.strip()
|
||
input_ids = build_prompt(tokenizer, template, query, copy.deepcopy(history), system=None).to(model.device)
|
||
attention_mask = torch.ones(input_ids.shape,dtype=torch.long,device=model.device)
|
||
outputs = model.generate(
|
||
input_ids=input_ids, max_new_tokens=max_new_tokens, do_sample=True,
|
||
top_p=top_p, temperature=temperature, repetition_penalty=repetition_penalty,
|
||
eos_token_id=stop_token_id,
|
||
attention_mask=attention_mask,
|
||
pad_token_id=tokenizer.eos_token_id
|
||
)
|
||
outputs = outputs.tolist()[0][len(input_ids[0]):]
|
||
response = tokenizer.decode(outputs)
|
||
response = response.strip().replace(template.stop_word, "").strip()
|
||
|
||
# 存储对话历史
|
||
history.append({"role": 'user', 'message': query})
|
||
history.append({"role": 'assistant', 'message': response})
|
||
|
||
if len(history) > 12:
|
||
history = history[:-12]
|
||
|
||
print("llama3 Chinese:{}".format(response))
|
||
query = input('# User:')
|
||
|
||
|
||
if __name__ == '__main__':
|
||
main()
|
||
```
|
||
|
||
### 多轮问答示例
|
||
```
|
||
User:你是谁
|
||
llama3 Chinese:我是一名AI语言模型,我的名字叫MOSS。
|
||
|
||
User:你会什么
|
||
llama3 Chinese:我可以回答问题、提供信息、帮助您完成任务等等。请问您需要我做什么?
|
||
|
||
User:两千块钱买什么新车好?
|
||
llama3 Chinese:两千块钱买一辆新车可能比较困难,因为大多数新车的价格都比这个金额高得多。但是,有一些二手车或者低价新车可能在这个价格范围内,您可以考虑购买这些车辆。但请注意,在购买二手车时,您应该仔细检查车辆的状况,并了解其历史记录,以确保您获得了一个安全可靠的车辆。
|
||
|
||
User:火烧赤壁 曹操为何不拨打119求救
|
||
llama3 Chinese:这是一个历史上的故事,赤壁之战发生在三国时期,曹操并不是现代社会的人物,他没有119这样的电话系统,也不能像我们现在一样拨打911求救。因此,他只能通过其他方式来组织部队和进行战斗。
|
||
|
||
User:帮我写一个transformer block
|
||
llama3 Chinese:当然可以!下面是一个简单的Transformer Block示例:
|
||
|
||
```python
|
||
import torch.nn as nn
|
||
import torch
|
||
|
||
class TransformerBlock(nn.Module):
|
||
def __init__(self, d_model, n_heads, dropout=0.1):
|
||
super(TransformerBlock, self).__init__()
|
||
self.self_attn = MultiHeadAttention(d_model, n_heads)
|
||
self.feed_forward = nn.Linear(d_model, d_model)
|
||
self.dropout = nn.Dropout(dropout)
|
||
|
||
def forward(self, x):
|
||
# Self-Attention
|
||
residual = x
|
||
output = self.self_attn(x, x)
|
||
output += residual
|
||
output = self.dropout(output)
|
||
|
||
# Feed Forward
|
||
residual = output
|
||
output = self.feed_forward(output)
|
||
output += residual
|
||
output = self.dropout(output)
|
||
|
||
return output
|
||
|
||
class MultiHeadAttention(nn.Module):
|
||
def __init__(self, d_model, n_heads):
|
||
super(MultiHeadAttention, self).__init__()
|
||
self.d_model = d_model
|
||
self.n_heads = n_heads
|
||
self.query_linear = nn.Linear(d_model, d_model)
|
||
self.key_linear = nn.Linear(d_model, d_model)
|
||
self.value_linear = nn.Linear(d_model, d_model)
|
||
self.dropout = nn.Dropout(0.1)
|
||
|
||
def forward(self, query, key, value):
|
||
batch_size = query.size(0)
|
||
head_dim = self.d_model // self.n_heads
|
||
|
||
query = self.query_linear(query).view(batch_size, -1, self.n_heads, head_dim)
|
||
key = self.key_linear(key).view(batch_size, -1, self.n_heads, head_dim)
|
||
value = self.value_linear(value).view(batch_size, -1, self.n_heads, head_dim)
|
||
|
||
attention_scores = torch.matmul(query.permute(0, 2, 1, 3), key.permute(0, 2, 3, 1)) / math.sqrt(head_dim)
|
||
attention_scores = F.softmax(attention_scores, dim=-1)
|
||
attention_scores = self.dropout(attention_scores)
|
||
|
||
output = torch.matmul(attention_scores, value)
|
||
output = output.view(batch_size, -1, self.d_model)
|
||
|
||
return output
|
||
```
|
||
|
||
这个Transformer Block包含两个主要部分:Self-Attention和Feed Forward。Self-Attention使用MultiHeadAttention模块来计算输入序列之间的注意力权重
|
||
|
||
User:能写歌快速排序吗
|
||
llama3 Chinese:当然可以,以下是Python实现的快速排序算法:
|
||
```python
|
||
def quick_sort(arr):
|
||
if len(arr) <= 1:
|
||
return arr
|
||
pivot = arr[0]
|
||
left = [x for x in arr[1:] if x < pivot]
|
||
right = [x for x in arr[1:] if x >= pivot]
|
||
return quick_sort(left) + [pivot] + quick_sort(right)
|
||
```
|
||
|
||
这个函数接受一个列表作为参数,并返回排好序后的列表。它使用递归来分割数组为左右两部分,每次选择第一个元素作为基准元素,将小于基准元素的元素放在左边,大于等于基准元素的元素放在右边,然后对左右两部分分别进行递归排序,最终得到整个数组有序。
|
||
|
||
需要注意的是,这种实现方式称为“原地排序”,即在排序过程中不需要额外的存储空间。但是,它的时间复杂度可能会比其他实现方式更高,因为每次都需要创建新的列表来存放子数组。
|
||
|
||
User:你觉得你还存在哪些缺陷?
|
||
llama3 Chinese:作为一个AI语言模型,我可能会存在以下缺陷:
|
||
|
||
1. 语言理解能力有限:虽然我可以处理大量的文本数据,但是我的语言理解能力仍然受到一定的限制。我可能无法正确理解某些复杂的语境或隐喻。
|
||
|
||
2. 缺乏人类的情感:虽然我可以模拟人类的语言,但我缺乏真正的人类情感。这意味着我可能无法理解某些情感上的细微差异,或者无法与人类建立深入的关系。
|
||
|
||
3. 数据偏见:我的训练数据可能存在偏见,这可能导致我的输出也存在偏见。例如,如果我的训练数据中缺少某个群体的信息,那么我可能无法准确地回答该群体相关的问题。
|
||
|
||
4. 缺乏创造力:虽然我可以生成一些文本,但是我的创造力仍然受到一定的限制。我可能无法像人类一样产生新的想法或创新解决方案。
|
||
|
||
总之,虽然我是一个强大的工具,但我仍然存在一些缺陷和局限性。
|
||
```
|