11

config.py

@@ -16,6 +16,12 @@ class Config:
     SWANLAB_PROJECT = "qweb3-sft-medical-10-11-1"
     # 验证用system的提示词
     PROMPT = "你是一个医学专家，你需要根据用户的问题，给出带有思考的回答。"
+    # 数据集question列名称
+    QUES_LABEL = "question"
+    # 数据集think列名称(可为空)
+    THINK_LABEL = "think"
+    # 数据集answer列名称
+    ANS_LABEL = "answer"
     DATA_MAX_LENGTH = 2048
 
 

dataset/delicate_medical_r1_data/default/train/downloads/01b7be597488efa04be3c3838a6cb36d08f9b1e99874a112d52790dd4050ad78.json

@@ -0,0 +1 @@
+{"url": "https://www.modelscope.cn/api/v1/datasets/krisfu/delicate_medical_r1_data/repo?Source=SDK&Revision=master&FilePath=r1_data_example.jsonl", "etag": null}

dataset/delicate_medical_r1_data/default/train/krisfu___delicate_medical_r1_data/default-f4a1046f7829f5d6/0.0.0/master/dataset_info.json

@@ -0,0 +1 @@
+{"description": "", "citation": "", "homepage": "", "license": "", "features": {"instruction": {"dtype": "string", "_type": "Value"}, "question": {"dtype": "string", "_type": "Value"}, "think": {"dtype": "string", "_type": "Value"}, "answer": {"dtype": "string", "_type": "Value"}, "metrics": {"quality_f1": {"dtype": "float64", "_type": "Value"}}}, "builder_name": "json", "dataset_name": "delicate_medical_r1_data", "config_name": "default", "version": {"version_str": "0.0.0", "major": 0, "minor": 0, "patch": 0}, "splits": {"train": {"name": "train", "num_bytes": 8978405, "num_examples": 2407, "dataset_name": "delicate_medical_r1_data"}}, "download_checksums": {"hf://datasets/krisfu/delicate_medical_r1_data@master/r1_data_example.jsonl": {"num_bytes": 9198097, "checksum": null}}, "download_size": 9198097, "dataset_size": 8978405, "size_in_bytes": 18176502}

eval_with_lora.py

@@ -0,0 +1,29 @@
+from transformers import AutoModelForCausalLM
+from peft import PeftModel, PeftConfig
+from config import Config,Default,Dir
+from modelscope import snapshot_download, AutoTokenizer
+import torch
+def load_model_and_tokenizer():
+    print("🧼 开始加载模型...")
+    # 在modelscope上下载Qwen模型到本地目录下
+    model_dir = snapshot_download(Config.MODEL_NAME)
+    # Transformers加载模型权重
+    tokenizer = AutoTokenizer.from_pretrained(Config.MODEL_NAME, use_fast=False,
+                                              trust_remote_code=True)
+    model = AutoModelForCausalLM.from_pretrained(model_dir, device_map="auto",
+                                                 torch_dtype=torch.bfloat16)
+    model.enable_input_require_grads()  # 开启梯度检查点时，要执行该方法
+    return model,tokenizer
+
+if __name__ == "__main__":
+    model, tokenizer = load_model_and_tokenizer()
+    # 2. 加载 LoRA 适配器配置和权重
+    peft_model_id = "./lora_adapter"  # 训练后保存的路径
+    peft_config = PeftConfig.from_pretrained(peft_model_id)
+    # 3. 将 LoRA 权重加载到原始模型（动态注入，不修改原始模型）
+    model_with_lora = PeftModel.from_pretrained(model, peft_model_id)
+    # 4. 进行推理（示例：文本生成）
+    input_text = "医生，我在研究内耳的前庭部分时，发现了一些特殊的结构，比如前庭嵴。请问前庭内还有哪些特殊的结构，它们的作用是什么？"
+    inputs = tokenizer(input_text, return_tensors="pt").to("cuda")
+    outputs = model.generate(**inputs, max_new_tokens=100)
+    print(tokenizer.decode(outputs[0], skip_special_tokens=True))

requirements.txt

@@ -3,7 +3,9 @@ modelscope
 yapf===0.32.0
 swanlab
 accelerate
+peft
 #pip install --upgrade modelscope -i https://mirrors.aliyun.com/pypi/simple/
 #pip install yapf==0.32.0 -i https://mirrors.aliyun.com/pypi/simple/
 #pip install swanlab -i https://mirrors.aliyun.com/pypi/simple/
-#pip install accelerate -i https://mirrors.aliyun.com/pypi/simple/
+#pip install accelerate -i https://mirrors.aliyun.com/pypi/simple/
+#pip install peft -i https://mirrors.aliyun.com/pypi/simple/

train.py

@@ -10,6 +10,62 @@ import json
 import random
 from config import Config,Default,Dir
 
+class Config:
+    # 基座模型
+    MODEL_NAME = "Qwen/Qwen3-0.6B"
+    # 数据集
+    DATASET_NAME = "krisfu/delicate_medical_r1_data"
+    # 数据集主题（子数据集）
+    DATASET_SUBJECT = "default"
+    # 数据集用途
+    DATASET_SPLIT = "train"
+    # 是否使用缓存
+    DATASET_USE_CACHE = True
+    # swanlab项目名称
+    SWANLAB_PROJECT = "qweb3-sft-medical-10-11-1"
+    # 验证用system的提示词
+    PROMPT = "你是一个医学专家，你需要根据用户的问题，给出带有思考的回答。"
+    # 数据集question列名称
+    QUES_LABEL = "question"
+    # 数据集think列名称(可为空)
+    THINK_LABEL = "think"
+    # 数据集answer列名称
+    ANS_LABEL = "answer"
+    DATA_MAX_LENGTH = 2048
+
+
+class Default:
+    DATASET_PATH = os.getenv("DATASET_PATH", "./dataset")  # 支持环境变量覆
+    MODEL_DATASET_PATH = os.getenv("MODEL_DATASET_PATH", "./model_dataset")  # 支持环境变量覆
+    SAVE_DIR = "./saved_model"  # 微调后模型存储位置
+    TRAIN_DATASET_FILE = "train.jsonl"
+    TEST_DATASET_FILE = "val.jsonl"
+    TRAIN_JSONL_NEW_FILE = "train_format.jsonl"
+    TEST_JSONL_NEW_FILE = "val_format.jsonl"
+
+
+dataset_short_name = Config.DATASET_NAME.split("/")[-1]
+model_dataset_short_name = Config.MODEL_NAME.split("/")[-1]
+# 确保缓存目录存在
+dataset_dir = os.path.normpath(
+        os.path.join(Default.DATASET_PATH, dataset_short_name, Config.DATASET_SUBJECT, Config.DATASET_SPLIT)
+    )
+model_dataset_DIR = os.path.normpath(
+        os.path.join(Default.MODEL_DATASET_PATH, model_dataset_short_name, Config.DATASET_SUBJECT, Config.DATASET_SPLIT)
+    )
+model_dir = os.path.normpath(
+        os.path.join(Default.SAVE_DIR,model_dataset_short_name, dataset_short_name, Config.DATASET_SUBJECT, Config.DATASET_SPLIT)
+    )
+os.makedirs(dataset_dir, exist_ok=True)
+os.makedirs(model_dataset_DIR, exist_ok=True)
+os.makedirs(model_dir, exist_ok=True)
+
+
+class Dir:
+    DATASET_DIR = dataset_dir
+    MODEL_DIR = model_dir
+    MODEL_DATASET_DIR = model_dataset_DIR
+
 
 
 
@@ -23,8 +79,13 @@ def dataset_jsonl_transfer(origin_path, new_path):
         for line in file:
             # 解析每一行的json数据
             data = json.loads(line)
-            input = data["question"]
+            input = data[Config.QUES_LABEL]
-            output = f"<think>{data['think']}</think> \n {data['answer']}"
+            output = ""
+            if hasattr(Config, "THINK_LABEL") and Config.THINK_LABEL in data:
+                output = data[Config.THINK_LABEL]
+            if hasattr(Config, "ANS_LABEL") and Config.ANS_LABEL in data:
+                output += f"{data[Config.ANS_LABEL]}"
+
             message = {
                 "instruction": Config.PROMPT,
                 "input": f"{input}",
@@ -173,7 +234,7 @@ def process_func(example):
     return {"input_ids": input_ids, "attention_mask": attention_mask, "labels": labels}
 
 
-def set_trainer(train_dataset):
+def set_trainer():
     print("🧼 正在生成训练对象...")
     print(f"🧼 {Dir.MODEL_DIR}")
     args = TrainingArguments(
@@ -191,6 +252,7 @@ def set_trainer(train_dataset):
         gradient_checkpointing=True,
         report_to="swanlab",
         run_name=Config.MODEL_NAME,
+        load_best_model_at_end=True,
     )
 
     trainer = Trainer(
@@ -204,6 +266,8 @@ def set_trainer(train_dataset):
 
 
 if __name__ == "__main__":
+    #释放显存
+    torch.cuda.empty_cache()
     #设置SWANLAB
     set_swanlab()
     #加载模型和分词器
@@ -212,26 +276,7 @@ if __name__ == "__main__":
     train_dataset, eval_dataset = load_train_and_eval_data()
     model_train_json_file, model_test_json_file = get_model_train_dataset_json_file()
     #设置训练对象
-    trainer = set_trainer(train_dataset)
+    trainer = set_trainer()
     print("🚬🚬🚬 开始模型训练...")
     trainer.train()
-    print(" 🎇🎇🎇模型训练完成，...")
+    print(" 🎇🎇🎇模型训练完成，...")
-    # 用测试集的前3条，主观看模型
-    test_df = pd.read_json(model_train_json_file, lines=True)[:3]
-    test_text_list = []
-    for index, row in test_df.iterrows():
-        instruction = row['instruction']
-        input_value = row['input']
-        messages = [
-            {"role": "system", "content": f"{instruction}"},
-            {"role": "user", "content": f"{input_value}"}
-        ]
-        response = predict(messages, model, tokenizer)
-        response_text = f"""
-        Question: {input_value}
-        LLM:{response}
-        """
-        test_text_list.append(swanlab.Text(response_text))
-        print(response_text)
-    swanlab.log({"Prediction": test_text_list})
-    swanlab.finish()

train_and_create_lora.py

@@ -0,0 +1,224 @@
+from modelscope import snapshot_download, AutoTokenizer
+from config import Config,Default,Dir
+from peft import LoraConfig, get_peft_model
+from transformers import AutoModelForCausalLM, TrainingArguments, Trainer, DataCollatorForSeq2Seq
+from modelscope.msdatasets import MsDataset
+import os
+import json
+import random
+import pandas as pd
+import torch
+from datasets import Dataset
+import swanlab
+
+lora_config = LoraConfig(
+    r=16,  # 低秩矩阵的秩（控制参数量，通常4-64）
+    lora_alpha=32,  # 缩放因子（影响更新强度）
+    target_modules=["q_proj", "v_proj"],  # 微调注意力层的Q/V矩阵（关键层）
+    lora_dropout=0.1,  # 防止过拟合
+    bias="none",  # 不微调偏置项
+    task_type="CAUSAL_LM",  # 任务类型（因果语言模型）
+)
+
+
+def load_model_and_tokenizer():
+    print("🧼 开始加载模型...")
+    # 在modelscope上下载Qwen模型到本地目录下
+    model_dir = snapshot_download(Config.MODEL_NAME)
+    # Transformers加载模型权重
+    tokenizer = AutoTokenizer.from_pretrained(Config.MODEL_NAME, use_fast=False,
+                                              trust_remote_code=True)
+    model = AutoModelForCausalLM.from_pretrained(model_dir, device_map="auto",
+                                                 torch_dtype=torch.bfloat16)
+    model.enable_input_require_grads()  # 开启梯度检查点时，要执行该方法
+    return model,tokenizer
+
+def mk_new_dataset_json(ds,dataset_train_json_file,dataset_test_json_file):
+    data_list = list(ds)
+    random.shuffle(data_list)
+    split_idx = int(len(data_list) * 0.9)
+    train_data = data_list[:split_idx]
+    val_data = data_list[split_idx:]
+    with open(dataset_train_json_file, 'w', encoding='utf-8') as f:
+        for item in train_data:
+            json.dump(item, f, ensure_ascii=False)
+            f.write('\n')
+    with open(dataset_test_json_file, 'w', encoding='utf-8') as f:
+        for item in val_data:
+            json.dump(item, f, ensure_ascii=False)
+            f.write('\n')
+    print(f"✅ 原始数据清洗完成 | 训练样本量: {len(train_data)} | 测试样本量: {len(train_data)}")
+
+def get_dataset_json():
+    # 确保缓存目录存在
+    os.makedirs(Dir.DATASET_DIR, exist_ok=True)
+    ds = MsDataset.load(
+        dataset_name=Config.DATASET_NAME,
+        subset_name=Config.DATASET_SUBJECT,
+        split=Config.DATASET_SPLIT,
+        cache_dir=Dir.DATASET_DIR  # 统一使用配置的路径
+    )
+    dataset_train_json_file = Dir.DATASET_DIR+"/"+Default.TRAIN_DATASET_FILE
+    dataset_test_json_file = Dir.DATASET_DIR + "/" + Default.TEST_DATASET_FILE
+    print(f"🏷️ 原始训练数据文件：{dataset_train_json_file}")
+    print(f"🏷️ 原始测试数据文件：{dataset_test_json_file}")
+    if (not os.path.exists(dataset_train_json_file)
+            or not os.path.exists(dataset_test_json_file)
+            or not Config.DATASET_USE_CACHE):
+        print("🏷️ 正在生成原始数据集...")
+        mk_new_dataset_json(ds,dataset_train_json_file,dataset_test_json_file)
+    return dataset_train_json_file,dataset_test_json_file
+
+def get_model_train_dataset_json_file():
+    model_train_json_file = Dir.MODEL_DATASET_DIR + "/" + Default.TRAIN_JSONL_NEW_FILE
+    model_test_json_file = Dir.MODEL_DATASET_DIR + "/" + Default.TEST_JSONL_NEW_FILE
+    return model_train_json_file, model_test_json_file
+def get_model_train_dataset_json():
+    model_train_json_file,model_test_json_file = get_model_train_dataset_json_file()
+    print(f"🏷️ 模型训练数据文件：{model_train_json_file}")
+    print(f"🏷️ 模型测试数据文件：{model_test_json_file}")
+    if not os.path.exists(model_train_json_file) or not os.path.exists(
+            model_test_json_file) or not Config.DATASET_USE_CACHE:
+        print("🏷️ 未找到模型对应数据集，准备从原始数据集进行生成数据集...")
+        dataset_train_json_file,dataset_test_json_file = get_dataset_json()
+        print("🏷️ 原始数据集生成成功...")
+        print("🧼 开始转换模型训练数据集...")
+        dataset_jsonl_transfer(dataset_train_json_file, model_train_json_file)
+        print("🧼 开始转换模型验证数据集...")
+        dataset_jsonl_transfer(dataset_test_json_file, model_test_json_file)
+    return model_train_json_file,model_test_json_file
+
+def dataset_jsonl_transfer(origin_path, new_path):
+    """
+    将原始数据集转换为大模型微调所需数据格式的新数据集
+    """
+    messages = []
+    # 读取旧的JSONL文件
+    with open(origin_path, "r") as file:
+        for line in file:
+            # 解析每一行的json数据
+            data = json.loads(line)
+            input = data[Config.QUES_LABEL]
+            output = ""
+            if hasattr(Config, "THINK_LABEL") and Config.THINK_LABEL in data:
+                output = data[Config.THINK_LABEL]
+            if hasattr(Config, "ANS_LABEL") and Config.ANS_LABEL in data:
+                output += f"{data[Config.ANS_LABEL]}"
+
+            message = {
+                "instruction": Config.PROMPT,
+                "input": f"{input}",
+                "output": output,
+            }
+            messages.append(message)
+    # 保存重构后的JSONL文件
+    with open(new_path, "w", encoding="utf-8") as file:
+        for message in messages:
+            file.write(json.dumps(message, ensure_ascii=False) + "\n")
+
+
+def predict(messages, model, tokenizer):
+    device = "cuda"
+    text = tokenizer.apply_chat_template(
+        messages,
+        tokenize=False,
+        add_generation_prompt=True
+    )
+    model_inputs = tokenizer([text], return_tensors="pt").to(device)
+    generated_ids = model.generate(
+        model_inputs.input_ids,
+        max_new_tokens=Config.DATA_MAX_LENGTH,
+    )
+    generated_ids = [
+        output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
+    ]
+    response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
+    return response
+
+def load_train_and_eval_data():
+    #获取训练数据集json.
+    print("🧼 开始获取训练数据...")
+    model_train_json_file,model_test_json_file = get_model_train_dataset_json()
+    # 得到训练集
+    train_df = pd.read_json(model_train_json_file, lines=True)
+    train_ds = Dataset.from_pandas(train_df)
+    train_dataset = train_ds.map(process_func, remove_columns=train_ds.column_names)
+    # 得到验证集
+    eval_df = pd.read_json(model_test_json_file, lines=True)
+    eval_ds = Dataset.from_pandas(eval_df)
+    eval_dataset = eval_ds.map(process_func, remove_columns=eval_ds.column_names)
+    return train_dataset,eval_dataset
+
+def process_func(example):
+    """
+    将数据集进行预处理
+    """
+    input_ids, attention_mask, labels = [], [], []
+    instruction = tokenizer(
+        f"<|im_start|>system\n{Config.PROMPT}<|im_end|>\n<|im_start|>user\n{example['input']}<|im_end|>\n<|im_start|>assistant\n",
+        add_special_tokens=False,
+    )
+    response = tokenizer(f"{example['output']}", add_special_tokens=False)
+    input_ids = instruction["input_ids"] + response["input_ids"] + [tokenizer.pad_token_id]
+    attention_mask = (
+        instruction["attention_mask"] + response["attention_mask"] + [1]
+    )
+    labels = [-100] * len(instruction["input_ids"]) + response["input_ids"] + [tokenizer.pad_token_id]
+    if len(input_ids) > Config.DATA_MAX_LENGTH:  # 做一个截断
+        input_ids = input_ids[:Config.DATA_MAX_LENGTH]
+        attention_mask = attention_mask[:Config.DATA_MAX_LENGTH]
+        labels = labels[:Config.DATA_MAX_LENGTH]
+    return {"input_ids": input_ids, "attention_mask": attention_mask, "labels": labels}
+
+
+def set_trainer():
+    print("🧼 正在生成训练对象...")
+    print(f"🧼 {Dir.MODEL_DIR}")
+    args = TrainingArguments(
+        output_dir=Dir.MODEL_DIR,
+        per_device_train_batch_size=1,
+        per_device_eval_batch_size=1,
+        gradient_accumulation_steps=4,
+        eval_strategy="steps",
+        eval_steps=500,
+        logging_steps=10,
+        num_train_epochs=2,
+        save_steps=500,
+        learning_rate=1e-4,
+        save_on_each_node=True,
+        gradient_checkpointing=True,
+        report_to="swanlab",
+        run_name=Config.MODEL_NAME,
+        load_best_model_at_end=True,
+    )
+
+    trainer = Trainer(
+        model=new_model,
+        args=args,
+        train_dataset=train_dataset,
+        eval_dataset=eval_dataset,
+        data_collator=DataCollatorForSeq2Seq(tokenizer=tokenizer, padding=True),
+    )
+    return trainer
+
+
+def set_swanlab():
+    os.environ["SWANLAB_PROJECT"] = Config.SWANLAB_PROJECT
+    swanlab.config.update({
+        "model": "",
+        "prompt": Config.PROMPT,
+        "data_max_length": Config.DATA_MAX_LENGTH,
+    })
+
+if __name__ == "__main__":
+    # 释放显存
+    torch.cuda.empty_cache()
+    # 设置SWANLAB
+    set_swanlab()
+    model,tokenizer = load_model_and_tokenizer()
+    train_dataset, eval_dataset = load_train_and_eval_data()
+    new_model = get_peft_model(model,lora_config)
+    trainer = set_trainer()
+    trainer.train()
+    new_model.save_pretrained("./lora_adapter")
+    tokenizer.save_pretrained("./lora_adapter")