minicpm-v 模型进行微调并进行验证
训练使用混合数据集进行训练,对minicpm-V进行lora微调,微调后使用llama3_1对输出结果与标签值进行比对,计算准确率。
验证代码为:
python
# URL = https://swift.readthedocs.io/zh-cn/latest/LLM/VLLM%E6%8E%A8%E7%90%86%E5%8A%A0%E9%80%9F%E4%B8%8E%E9%83%A8%E7%BD%B2.html
'''
使用swift进行部署的示例
model: llama3_1-8b-instruct
CUDA_VISIBLE_DEVICES=2 swift deploy --max_model_len 4096 --model_type llama3_1-8b-instruct --model_id_or_path /nas/share/model/huggingface/models--meta-llama--Meta-Llama-3.1-8B-Instruct/snapshots/8c22764a7e3675c50d4c7c9a4edb474456022b16
'''
# 客户端
# example of using swift client
import os
import json
from tqdm import tqdm
from swift.llm import get_model_list_client, XRequestConfig, inference_client
model_list = get_model_list_client()
model_type = model_list.data[0].id
def get_data_jsonl(data_path):
datas =[]
with open(data_path, 'r') as f:
data = f.readlines()
for i in range(len(data)):
datas.append(json.loads(data[i]))
return datas
def save_data_jsonl(save_path:str, datas: list):
with open(save_path, 'w') as f:
for data in datas:
json.dump(data, f,ensure_ascii=False)
f.write('\n')
return save_path
def get_save_path(data_path:str):
save_path =os.path.splitext(data_path)[0] + '_llm_eval.jsonl'
if not os.path.exists(os.path.dirname(save_path)):
os.makedirs(os.path.dirname(save_path))
return save_path
def get_result_correct_path(data_path:str):
result_correct_path =os.path.splitext(data_path)[0] + '_llm_eval_correct.json'
if not os.path.exists(os.path.dirname(result_correct_path)):
os.makedirs(os.path.dirname(result_correct_path))
return result_correct_path
def save_result_correct(data_path:str,precision:float):
result = {
'data_path': data_path,
'precision': precision
}
with open(data_path, 'w') as f:
json.dump(result, f,ensure_ascii=False)
print(f'precision: {precision*100}%')
return result
def main():
def q_tempalte(response, reply):
question_template = '''你可以作为一个语言专家,判断一下两个回答是否相同吗 \n
"response": "{}",
"reply": "{}",
----
如果是,返回"YES",否则返回"NO"。
'''.format(response, reply)
return question_template
# origin
data_paths =
for data_path in tqdm(data_paths):
datas = get_data_jsonl(data_path)
save_data =[]
total_correct = 0
total = 0
for data in tqdm(datas):
response = data['response']
reply = data['reply']
question = q_tempalte(response, reply)
request_config = XRequestConfig(max_tokens=32, temperature=0.1, seed=42)
resp = inference_client(model_type, question, request_config=request_config)
response = resp.choices[0].message.content
data['llm_response'] = response
save_data.append(data)
if response.lower() == 'yes':
total_correct += 1
total += 1
precision = total_correct / total
# 保存数据
save_path = get_save_path(data_path)
save_data_jsonl(save_path, save_data)
# 保存精度
result_correct_path = get_result_correct_path(data_path)
save_result_correct(result_correct_path, precision)
if __name__ == '__main__':
main()