Gemma

1.使用

首先是去HF下载模型，但一直下载不了，所以去了HF镜像网站，下载gemma需要HF的Token，按照步骤就可以下载。代码主要是Kaggle论坛里面的分享内容。

huggingface-cli download --token hf_XXX --resume-download google/gemma-7b --local-dir gemma-7b-mirror

这里我有时是2b有时是7b，换着用。

from transformers import AutoTokenizer, AutoModelForCausalLM  
tokenizer = AutoTokenizer.from_pretrained("D:/Gemma/gemma-2b-int-mirror2")
Gemma = AutoModelForCausalLM.from_pretrained("D:/Gemma/gemma-2b-int-mirror2")
def answer_the_question(question):input_ids = tokenizer(question, return_tensors="pt")generated_text = Gemma.generate(**input_ids,max_length=256)answer = tokenizer.decode(generated_text[0], skip_special_tokens=True)return answer
question = "给我写一首优美的诗歌?"
answer = answer_the_question(question)
print(answer)

2.RAG

参考

from langchain_community.document_loaders import PyPDFLoader
from langchain.text_splitter import CharacterTextSplitter
from langchain_community.embeddings import HuggingFaceEmbeddings
from langchain_community.vectorstores import FAISS
##2.1 根据question检索sentence chunk
import os
def get_all_pdfs(directory):pdf_files = []for root, dirs, files in os.walk(directory):for file in files:if file.endswith(".pdf"):pdf_files.append(os.path.join(root, file))return pdf_filesclass RAG:def __init__(self, num_retrieved_docs=5, pdf_folder_path='D:/Gemma/PDF'):pdf_files = get_all_pdfs(pdf_folder_path)print("Documents used", pdf_files)loaders = [PyPDFLoader(pdf_file) for pdf_file in pdf_files]all_documents = []for loader in loaders:raw_documents = loader.load()text_splitter = CharacterTextSplitter(separator="\n\n",chunk_size=10,chunk_overlap=1,# length_function=len,)documents = text_splitter.split_documents(raw_documents)all_documents.extend(documents)embeddings = HuggingFaceEmbeddings(model_name="D:/Projects/model/m3e-base")    self.db = FAISS.from_documents(all_documents, embeddings)self.retriever = self.db.as_retriever(search_kwargs={"k": num_retrieved_docs})def search(self, query):docs = self.retriever.get_relevant_documents(query)return docs
retriever = RAG()
##2.2根据sentence chunk和question去回答
class Assistant:def __init__(self):self.tokenizer = AutoTokenizer.from_pretrained("D:/Gemma/gemma-2b-int-mirror2")self.Gemma = AutoModelForCausalLM.from_pretrained("D:/Gemma/gemma-2b-int-mirror2")def create_prompt(self, query, retrieved_info):prompt = f"""你是人工智能助手，需要根据Relevant information里面的相关内容回答用户的Instruction，其中相关信息如下：Instruction: {query}Relevant information: {retrieved_info}Output:"""print(prompt)return promptdef reply(self, query, retrieved_info):prompt = self.create_prompt(query, retrieved_info)input_ids = self.tokenizer(query, return_tensors="pt").input_ids# Generate text with a focus on factual responsesgenerated_text = self.Gemma.generate(input_ids,do_sample=True,max_length=500,temperature=0.7, # Adjust temperature according to the task, for code generation it can be 0.9)# Decode and return the answeranswer = self.tokenizer.decode(generated_text[0], skip_special_tokens=True)return answer
chatbot = Assistant()
## 2.3开始使用RAG
def generate_reply(query):related_docs = retriever.search(query)#print('related docs', related_docs)reply = chatbot.reply(query, related_docs)return reply
reply = generate_reply("存在的不足及后续的优化工作")
for s in reply.split('\n'):print(s)

3.LoRA

3.1LoRA分类任务

参考
使用nlp-getting-started数据集训练模型做二分类任务。首先拿到源model

from datasets import load_dataset
from transformers import AutoTokenizer,AutoModelForSequenceClassification, DataCollatorWithPadding, Trainer, TrainingArguments,pipeline
from peft import prepare_model_for_int8_training,LoraConfig, TaskType, get_peft_model
import numpy as np
NUM_CLASSES = 2#模型输出分类的类别数
BATCH_SIZE,EPOCHS,R,LORA_ALPHA,LORA_DROPOUT = 8,5,64,32,0.1#LoRA训练的参数
MODEL_PATH="D:/Gemma/gemma-2b-int-mirror2"#模型地址
# 1.源model，设置输出二分类
tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH)
model = AutoModelForSequenceClassification.from_pretrained(MODEL_PATH,num_labels=NUM_CLASSES)
print(model)

处理csv数据，将输入文字经过tokenizer编码处理

#2.处理dataset，输入过长进行truncation（tokenizer处理后）
dataset = load_dataset('csv', data_files='D:/Gemma/nlp-getting-started/train.csv')
dataset['test'] = dataset['train']
dataset = dataset.remove_columns(['id', 'keyword', 'location'])
dataset = dataset.rename_column("target", "label")#csv最后只保留了text列和label列
tokenized_dataset = {}#train和test
for split in dataset.keys():tokenized_dataset[split] = dataset[split].map(lambda x: tokenizer(x["text"], truncation=True), batched=True)
print(tokenized_dataset["train"])
print(tokenized_dataset["train"][1])

在源model基础上配置LoRA的参数，形成lora_model

#3.LoRA模型参数设置
model = prepare_model_for_int8_training(model)
lora_config = LoraConfig(r=R,lora_alpha=LORA_ALPHA,lora_dropout=LORA_DROPOUT,task_type=TaskType.SEQ_CLS,#SEQ_CLS:序列分类任务；TOKEN_CLS命名实体识别；SEQ2SEQ机器翻译；LM语言建模任务target_modules='all-linear'#all-linear所有线性层；embeddings嵌入层；convs卷积层
)
lora_model = get_peft_model(model, lora_config)
print(lora_model)
print(lora_model.print_trainable_parameters())#LoRA模型要训练的参数

配置lora_model的训练参数

#4.LoRA训练参数设置（损失计算等）
def compute_metrics(eval_pred):predictions, labels = eval_predpredictions = np.argmax(predictions, axis=1)return {"accuracy": (predictions == labels).mean()}trainer = Trainer(model=lora_model,args=TrainingArguments(output_dir="./LoAR_data/",learning_rate=2e-5,per_device_train_batch_size=BATCH_SIZE,per_device_eval_batch_size=BATCH_SIZE,evaluation_strategy="epoch",save_strategy="epoch",num_train_epochs=EPOCHS,weight_decay=0.01,load_best_model_at_end=True,logging_steps=10,report_to="none"),train_dataset=tokenized_dataset["train"],eval_dataset=tokenized_dataset["test"],tokenizer=tokenizer,data_collator=DataCollatorWithPadding(tokenizer=tokenizer),compute_metrics=compute_metrics,
)

开始训练并保存使用模型

#5.训练并评估
print("Evaluating the Model Before Training!")
trainer.evaluate()
print("Training the Model")
trainer.train()
print("Evaluating the trained model")
trainer.evaluate()
#6.保存并使用
lora_model.save_pretrained('fine-tuned-model')
clf = pipeline("text-classification", lora_model, tokenizer=MODEL_PATH)#LoRA训练后的模型

3.2LoRA中文建模任务

参考
首先拿到源model和config

from transformers import AutoConfig,AutoTokenizer,AutoModelForCausalLM, Trainer, TrainingArguments
from peft import LoraConfig, get_peft_model,prepare_model_for_kbit_training,PeftModel
import torch
import  datasets
from tqdm import tqdm
import json
BATCH_SIZE,EPOCHS,R,LORA_ALPHA,LORA_DROPOUT = 8,5,64,32,0.1#LoRA训练的参数
MODEL_PATH="D:/Gemma/gemma-2b-int-mirror2"#模型地址
device = torch.device('cuda:0')
# 1.源model和model的config
config = AutoConfig.from_pretrained(MODEL_PATH, trust_remote_code=True)
config.is_causal = True  #确保模型在生成文本时只能看到左侧的上下文
tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(MODEL_PATH,device_map="auto", config=config,trust_remote_code=True)

根据模型和config处理json数据

#2.根据model的config处理dataset（tokenizer处理后），并保存加载
def preprocess(tokenizer: PreTrainedTokenizer, config, file_path, max_seq_length, prompt_key, target_key, skip_overlength=False):  # 数据预处理  pad_token_id = tokenizer.pad_token_id  # 获取填充标记的ID  with open(file_path, "r", encoding="utf8") as f:  for line in tqdm(f.readlines()):  example = json.loads(line)  prompt_ids = tokenizer.encode(example[prompt_key], max_length=max_seq_length, truncation=True)  target_ids = tokenizer.encode(example[target_key], max_length=max_seq_length, truncation=True)  # 检查prompt和target连接后是否超出最大长度，并在需要时跳过  total_length = len(prompt_ids) + len(target_ids) + (1 if config.eos_token_id is not None else 0)  if skip_overlength and total_length > max_seq_length:  continue  # 连接prompt和target，并添加EOS标记（如果提供）  input_ids = prompt_ids + target_ids  if config.eos_token_id is not None:  input_ids.append(config.eos_token_id)  # 截断序列到最大长度  input_ids = input_ids[:max_seq_length]  # 填充序列到最大长度  input_ids.extend([pad_token_id] * (max_seq_length - len(input_ids)))  assert len(input_ids) == max_seq_length, "序列长度必须等于max_seq_length"  yield {  "input_ids": input_ids,  "seq_len": len(prompt_ids)  # 注意：这里提供的seq_len是原始prompt的长度，不包括填充  }
dataset = datasets.Dataset.from_generator(lambda: preprocess(tokenizer, config, "D:/Gemma/try/hc3_chatgpt_zh_specific_qa.json", max_seq_length=2000, prompt_key="q",target_key="a",))dataset.save_to_disk("h3c-chinese")  # 保存处理后的数据集
train_set = datasets.load_from_disk("h3c-chinese")#加载处理后的数据集

配置Lora参数

#3.LoRA模型参数设置
model = prepare_model_for_kbit_training(model)
lora_config = LoraConfig(r=R,lora_alpha=LORA_ALPHA,lora_dropout=LORA_DROPOUT,task_type="CAUSAL_LM",target_modules='all-linear'
)
lora_model = get_peft_model(model, lora_config)
print(lora_model)
print(lora_model.print_trainable_parameters())#LoRA模型要训练的参数

配置lora的训练参数，包括损失计算compute_metrics，并对输入的input_ids构造输入样本列表批次处理。

trainer = Trainer(model=lora_model,args=TrainingArguments(output_dir="./LoAR_data2/",learning_rate=2e-5,per_device_train_batch_size=BATCH_SIZE,save_strategy="epoch",num_train_epochs=EPOCHS,weight_decay=0.01,logging_steps=10,report_to="none"),train_dataset=train_set,tokenizer=tokenizer,data_collator=DataCollatorWithPadding(tokenizer=tokenizer),
#     compute_metrics=compute_metrics
)
trainer.train()