1. Hard Prompt

定義： Hard prompt 是一種更為具體和明確的提示，要求模型按照給定的信息生成精確的結果，通常用于需要模型提供準確答案的任務.

原理： Prompt Tuning原理如下圖所示：凍結主模型全部參數，在訓練數據前加入一小段Prompt，只訓練Prompt的表示層，即一個Embedding模塊。論文實驗表明，只要模型規模夠大，簡單加入 Prompt tokens 進行微調，就能取得很好的效果。

優點： 資源消耗比 Soft Prompt 小，容易收斂。

2. Soft Prompt

**定義：**Soft prompt 通常指的是一種較為寬泛或模糊的提示，允許模型在生成結果時有更大的自由度，通常用于啟發模型進行創造性的生成。

缺點： 資源消耗大，參數都是隨機初始化的，Loss 會有震蕩。

from datasets import load_from_disk
from transformers import AutoTokenizer, AutoModelForCausalLM, DataCollatorForSeq2Seq
from transformers import pipeline, TrainingArguments, Trainer
from peft import PromptTuningConfig, get_peft_model, TaskType, PromptTuningInit, PeftModel# 分詞器
tokenizer = AutoTokenizer.from_pretrained("Langboat/bloom-1b4-zh")# 函數內將instruction和response拆開分詞的原因是：
# 為了便于mask掉不需要計算損失的labels, 即代碼labels = [-100] * len(instruction["input_ids"]) + response["input_ids"]
def process_func(example):MAX_LENGTH = 256input_ids, attention_mask, labels = [], [], []instruction = tokenizer("\n".join(["Human: " + example["instruction"], example["input"]]).strip() + "\n\nAssistant: ")response = tokenizer(example["output"] + tokenizer.eos_token)input_ids = instruction["input_ids"] + response["input_ids"]attention_mask = instruction["attention_mask"] + response["attention_mask"]labels = [-100] * len(instruction["input_ids"]) + response["input_ids"]if len(input_ids) > MAX_LENGTH:input_ids = input_ids[:MAX_LENGTH]attention_mask = attention_mask[:MAX_LENGTH]labels = labels[:MAX_LENGTH]return {"input_ids": input_ids,"attention_mask": attention_mask,"labels": labels}if __name__ == "__main__":# 加載數據集dataset = load_from_disk("./PEFT/data/alpaca_data_zh")# 處理數據tokenized_ds = dataset.map(process_func, remove_columns = dataset.column_names)# print(tokenizer.decode(tokenized_ds[1]["input_ids"]))# print(tokenizer.decode(list(filter(lambda x: x != -100, tokenized_ds[1]["labels"]))))# 創建模型model = AutoModelForCausalLM.from_pretrained("Langboat/bloom-1b4-zh", low_cpu_mem_usage=True)# 設置 Prompt-Tuning# Soft Prompt# config = PromptTuningConfig(task_type=TaskType.CAUSAL_LM, num_virtual_tokens=10) # soft_prompt會隨機初始化# Hard Promptconfig = PromptTuningConfig(task_type = TaskType.CAUSAL_LM,prompt_tuning_init = PromptTuningInit.TEXT,prompt_tuning_init_text = "下面是一段人與機器人的對話。", # 設置hard_prompt的具體內容num_virtual_tokens = len(tokenizer("下面是一段人與機器人的對話。")["input_ids"]),tokenizer_name_or_path = "Langboat/bloom-1b4-zh")model = get_peft_model(model, config) # 生成Prompt-Tuning對應的modelprint(model.print_trainable_parameters())# 訓練參數args = TrainingArguments(output_dir = "/tmp_1203",per_device_train_batch_size = 1,gradient_accumulation_steps = 8,logging_steps = 10,num_train_epochs = 1)# trainertrainer = Trainer(model = model,args = args,train_dataset = tokenized_ds,data_collator = DataCollatorForSeq2Seq(tokenizer = tokenizer, padding = True))# 訓練模型trainer.train()# 模型推理model = AutoModelForCausalLM.from_pretrained("Langboat/bloom-1b4-zh", low_cpu_mem_usage=True)peft_model = PeftModel.from_pretrained(model = model, model_id = "/tmp_1203/checkpoint-500/")peft_model = peft_model.cuda()ipt = tokenizer("Human: {}\n{}".format("考試有哪些技巧？", "").strip() + "\n\nAssistant: ", return_tensors="pt").to(peft_model.device)print(tokenizer.decode(peft_model.generate(**ipt, max_length=128, do_sample=True)[0], skip_special_tokens=True))