Alpaca(https://github.com/tatsu-lab/stanford_alpaca)在70亿参数的LLaMA-7B上进行微调,通过52k指令数据(https://github.com/tatsu-lab/stanford_alpaca/blob/main/alpaca_data.json),在8个80GB A100上训练3小时,在对话上的性能比肩GPT-3.5。
指令微调数据来源:self-instruct
52k数据所对应的alpaca_data.json是一个字典列表,每个字典包含以下字段:
- instruction:52k条指令中的每一条都是唯一的;
- input:可以是空的(以instruction作为上下文),也可以是与instruction相关的文本;
- output:由GPT3.5的API生成的对应回答。
比如在中文版52k指令(https://github.com/hikariming/alpaca_chinese_dataset)中:
[
{
"instruction": "\"给出三个保持健康的小贴士。\"",
"input": "",
"output": "1. 饮食要均衡,确保摄入足够的水果和蔬菜。\n2. 经常锻炼,保持身体活跃和强壮。\n3. 要保证充足的睡眠,并保持一个稳定的睡眠时间表。"
},
{
"instruction": "\"三原色是什么?\"",
"input": "",
"output": "\"三原色是红色、蓝色和黄色。\""
},
...
{
"instruction": "将以下内容分类为动物、植物和矿物。",
"input": "\"橡树,铜矿石,大象\"",
"output": "\"橡树:植物\n铜矿石:矿物\n大象:动物\""
},
]
关于52k数据,是通过openAI的API用self-instruct方式生成的,花费500美元,self-instruct是将预训练语言模型与指令对齐的方法(https://github.com/yizhongw/self-instruct):
一共包含4个步骤:
- Step1:通过模型生成新的指令,根据人工设计的175个任务,每个任务都有对应的(指令,输入,输出)或(指令,输出);使用模型生成新的指令;
- Step2:对模型生成的指令进行判断(指令是否是一个分类任务);
- Step3:根据Step2的判断结果,给出不同的输出,
如果是分类任务,就通过模型输出 Class_label 和 Input(Output-first);
如果不是分类任务,就通过模型输出 Input 和 Output(Input-first); - Step4:过滤及后处理,对上述模型生成的数据进行过滤和后处理,将经过过滤和后处理的数据添加到种子池中。
对于以上4个步骤进行不断循环,直到种子池有足够多的数据(通常会设定一个具体的参数,比如:52000),生成过程停止。
比如在项目 https://github.com/yizhongw/self-instruct 中,依次执行:
# 1. Generate instructions from the seed tasks
./scripts/generate_instructions.sh
# 2. Identify whether the instruction represents a classification task or not
./scripts/is_clf_or_not.sh
# 3. Generate instances for each instruction
./scripts/generate_instances.sh
# 4. Filtering, processing, and reformatting
./scripts/prepare_for_finetuning.sh
175个人工设计的种子任务在 https://github.com/yizhongw/self-instruct/blob/main/data/seed_tasks.jsonl 中:
{"id": "seed_task_0",
"name": "breakfast_suggestion",
"instruction": "Is there anything I can eat for a breakfast that doesn't include eggs, yet includes protein, and has roughly 700-1000 calories?",
"instances":
[{
"input": "",
"output": "Yes, you can have 1 oatmeal banana protein shake and 4 strips of bacon. The oatmeal banana protein shake may contain 1/2 cup oatmeal, 60 grams whey protein powder, 1/2 medium banana, 1tbsp flaxseed oil and 1/2 cup watter, totalling about 550 calories. The 4 strips of bacon contains about 200 calories."
}],
"is_classification": false}
...
{"id": "seed_task_174",
"name": "fact_checking",
"instruction": "Fact checking - tell me if the statement is true, false, or unknown, based on your knowledge and common sense.",
"instances":
[{
"input": "Philadelphia is among the top 10 safest cities in the US.",
"output": "false"
}],
"is_classification": true}
执行:
python self_instruct/bootstrap_instructions.py --batch_dir "自己定义的目录,比如data/gpt3.5" --num_instructions_to_generate 100 --seed_tasks_path data/seed_tasks.jsonl --engine "davinci" --api_key "自己的openai API"
上述指令生成100条数据,这只会产生较少费用,生成数据会写入data/gpt3.5/machine_generated_instructions.jsonl中,这些数据是通过openAI的API生成了与种子任务关联度比较弱的一些任务描述(因为相似度高的对微调没有用)。
然后判断是否为分类任务:
python self_instruct/identify_clf_or_not.py --batch_dir data/gpt3.5 --engine "davinci" --request_batch_size 5 --api_key "自己的openai API"
结果写入data/gpt3.5/is_clf_or_not_davinci_template_1.jsonl中,然后根据步骤2的结果生成输出:
python self_instruct/generate_instances.py --batch_dir data/gpt3.5 --input_file machine_generated_instructions.jsonl --output_file machine_generated_instances.jsonl --max_instances_to_gen 5 --engine "davinci" --request_batch_size 5 --api_key "自己的openai API"
结果写入 data/gpt3.5/machine_generated_instances.jsonl中,然后进行过滤和后处理:
python self_instruct/prepare_for_finetuning.py --instance_files data/gpt3.5/machine_generated_instances.jsonl --classification_type_files data/gpt3.5/is_clf_or_not_davinci_template_1.jsonl --output_dir data/gpt3.5/finetuning_data --include_seed_tasks --seed_tasks_path data/seed_tasks.jsonl
运行后会生成两个数据文件,均在data/gpt3.5/finetuning_data目录下:
- all_generated_instances.jsonl,all_generated_instances.jsonl中包含的是 instruction,input,output,这是用于微调LLaMA-7B的格式。
- gpt3_finetuning_data_xxx.jsonl,包含的是prompt,completion,这是用于微调GPT3的格式。
Alpaca-LoRA
LoRA可以降低微调LLM的成本,在神经⽹络模型中,模型参数通常以矩阵的形式表示。对于⼀个预训练好的模型,其参数矩阵已经包含了很多有⽤的信息。为了使模型适应特定任务,需要对这些参数进⾏微调。LoRA是一种思想:用低秩的方法调整参数矩阵,低秩表示一个矩阵可以用两个小矩阵相乘近似(LoRA: LOW-RANK ADAPTATION OF LARGE LANGUAGE MODELS)。
LoRA包含以下步骤:
- 1.选择目标层:首先,在预训练神经网络模型中选择要应用LoRA的目标层,这些层通常是与特定任务相关的,比如自注意力机制中的Q和K矩阵;
- 2.初始化映射矩阵和逆映射矩阵:为目标层创建两个较小的矩阵A和B;
A是映射矩阵,一般用随机高斯分布初始化,deepspeed chat中用LoRA策略时则通过0矩阵占位,A矩阵用于降维;
B是逆映射矩阵,用0矩阵初始化,用于升维; - 3.参数变换:将目标层的原始参数矩阵W通过A和B进行变换: W ′ = W + A B W'=W+AB W′=W+AB, W ′ W' W′是变换后的参数矩阵;
- 4.微调:使用 W ′ W' W′替换 W W W在特定任务的训练数据上进行微调;
- 5.梯度更新:在微调过程中,计算损失函数关于映射矩阵A和逆映射矩阵B的梯度,并使⽤优化算法,如Adam、SGD对A和B进⾏更新,注意,在更新过程中,原始参数矩阵W保持不变,即训练的时候固定原始LLM的参数,只训练A和B;
- 6.重复更新:重复步骤3-5,直到达到预定的epoch或模型收敛。
HuggingFace已经将LoRA封装到了PEFT中(Parameter-Efficient Fine-Tuning),PEFT库可以使预训练语⾔模型⾼效适应各种下游任务,⽽⽆需微调模型的所有参数,即仅微调少量模型参数,从⽽⼤⼤降低了计算和存储成本。
历史:
Alpaca率先带动self-instruct,启发后续的人也采用提示GPT API的方式生成数据,比如BELLE、ChatLLaMA、ColossalChat,从而解决数据扩展的问题。然后又有新的LLM用Alpaca去生成新的数据进行微调,⽐如ChatDoctor ⽤到Alpaca的数据进⾏微调,有⼈用BELLE数据微调chatGLM。
微调LLaMA-7B
下载Alpaca-LoRA项目,并安装所需的依赖:
$ git clone https://github.com/tloen/alpaca-lora.git
$ pip install -r requirements.txt
下载预训练模型的权重,以及斯坦福进一步清洗后的微调数据(原本的52k数据中存在一些有问题的信息):
$ git clone https://huggingface.co/decapoda-research/llama-7b-hf
$ git clone https://huggingface.co/datasets/yahma/alpaca-cleaned
预训练模型包含33个405MB的bin文件,大约占14GB内存。
在alpaca-lora-main/finetune.py中,设置batch_size=4(micro_batch_size: int = 4
)以适配16GB的单个GPU(显存占用9GB),由于微调时间很长,大约60h,所以新建finetune.sh后台运行:
nohup python -u finetune.py \
--base_model '/data/temp/my-alpaca-lora/llama-7b-hf' \
--data_path '/students/julyedu_636353/alpaca-lora-main/alpaca-cleaned' \
--output_dir '/data/temp/my-alpaca-lora' \
>> log.out 2>&1 & # 后台运行, 日志写到 log.out
可以直接获取已经训练好的LoRA权重(67MB):
git clone https://huggingface.co/tloen/alpaca-lora-7b
或者获取通过GPT4生成指令数据微调后的LoRA权重(模型为LLaMA-7B,主要微调方式为Alpaca,低成本的微调策略为LoRA),故称LoRA权重为适配器adapter weights,GPT4对应的LoRA权重也应该是67MB:
git clone https://huggingface.co/chansung/gpt4-alpaca-lora-7b
利用alpaca-lora-main/generate.py进行推理,其中使用import gradio as gr
实现了快捷的可视化界面,新建inference.sh,推理时占用显存8GB:
python generate.py \
--load_8bit \
--base_model '/data/temp/my-alpaca-lora/llama-7b-hf' \
--lora_weights 'home/user/alpaca-lora-main/gpt4-alpaca-lora-7b'
对于一个问题,单个GPU上生成所需时间依旧很慢,大约1分钟,示例如下:
生成回答为:
Based on the MRI scan of the patient’s brain, it is possible that the patient may have Alzheimer’s disease. However, it is important to note that the presence of a “false shadow” in the MRI scan caused by the patient’s physical activity does not necessarily mean that the patient has Alzheimer’s disease. There are several methods that can be used to reduce or eliminate the “false shadow” in the MRI scan, such as:
Asking the patient to remain still during the MRI scan.
Asking the patient to wear earplugs.
再尝试一个新问答:
生成的回答为:
The expression of Alzheimer’s disease is a decline in memory.
The doctor showed a scenario in which there was a dog and a person playing the violin in a garden. When the patient was asked to recall the scene, the patient did not mention the dog, which could indicate that the patient may have Alzheimer’s disease.
However, it is important to note that this is only one scenario and does not necessarily mean that the patient has Alzheimer’s disease. It is recommended that the patient be evaluated by a medical professional to confirm the diagnosis and receive appropriate treatment and care.
也可以只使用instruct进行问答:
生成的回答为:
Alzheimer’s disease is a progressive neurodegenerative disorder that affects memory, thinking, and behavior. It is the most common form of dementia, accounting for 60 to 80 percent of cases. The exact cause of Alzheimer’s is unknown, but it is believed to be the result of a combination of genetic, environmental, and lifestyle factors. There is no cure for Alzheimer’s, but medications and lifestyle changes can help manage symptoms and slow the progression of the disease.