构建AI智能体：AI古典文学：基于LoRA微调本地大模型打造唐诗生成器

import torch
from modelscope import AutoModelForCausalLM, AutoTokenizer, snapshot_download
from peft import LoraConfig, get_peft_model
from datasets import Dataset
import random

# 设置设备
device = "cpu"
print("使用CPU运行")

# 加载超小模型
model_id = "qwen/Qwen1.5-0.5B-Chat"
cache_dir = "D:\\modelscope\\hub"
print("正在下载/校验模型缓存...")
local_model_path = snapshot_download(model_id, cache_dir=cache_dir)

print("加载模型中...")

tokenizer = AutoTokenizer.from_pretrained(local_model_path, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    local_model_path,
    trust_remote_code=True,
    torch_dtype=torch.float32,
    device_map=device
)

if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

print("模型加载完成！")

def create_large_original_poetry_data():
    """创建大规模原创唐诗数据集"""
    
    training_data = []
    
    # 1. 四季主题原创诗 (8首)
    seasons_poems = [
        {
            "instruction": "创作一首描写春天桃花的原创唐诗",
            "output": "《春桃》\n春来桃李满园开，粉白相间映日来。\n蝶舞蜂喧花似海，香风阵阵入襟怀。\n游人络绎争相赏，墨客挥毫赋雅才。\n莫道芳华容易逝，且将春色用心裁。"
        },
        {
            "instruction": "写一首夏天荷塘的原创唐诗",
            "output": "《夏荷》\n炎炎夏日荷塘畔，翠盖红妆映日鲜。\n玉立亭亭尘不染，清香袅袅韵无边。\n蜻蜓点水波纹起，锦鲤嬉游戏叶间。\n愿效莲花高洁志，冰心一片向青天。"
        },
        {
            "instruction": "创作秋天枫叶的原创唐诗",
            "output": "《秋枫》\n秋深枫叶似火红，漫山遍野映苍穹。\n霜侵露染色愈艳，风拂枝摇韵更浓。\n游人驻足观奇景，墨客挥毫绘彩容。\n莫道秋光多寂寥，枫情如火暖胸中。"
        },
        {
            "instruction": "写冬天雪景的原创唐诗",
            "output": "《冬雪》\n寒冬凛冽雪花飞，万里山河着素衣。\n玉树琼枝添雅趣，银装素裹显神奇。\n儿童嬉戏堆雪偶，老叟围炉话旧时。\n莫道冬来无美景，冰清玉洁亦相宜。"
        },
        {
            "instruction": "创作早春的原创唐诗",
            "output": "《早春》\n残雪消融春意萌，寒梅吐艳报新晴。\n东风暗度千山绿，细雨悄润万物生。\n燕子初归寻旧垒，农夫早作备春耕。\n一年好景从今始，莫负光阴奋力行。"
        },
        {
            "instruction": "写盛夏雷雨的原创唐诗",
            "output": "《夏雷》\n炎炎盛夏闷如蒸，忽见乌云压古城。\n电闪雷鸣惊天地，风狂雨骤洗尘清。\n荷塘水满蛙声闹，竹径风凉暑气平。\n雨过天晴虹彩现，山河焕然一新明。"
        },
        {
            "instruction": "创作深秋的原创唐诗",
            "output": "《深秋》\n霜降秋深寒意浓，千山万木尽丹红。\n菊花开遍东篱下，雁阵南飞碧空中。\n稻谷丰收仓廪实，果香弥漫果园丰。\n莫悲秋老芳华逝，且看枫情胜春容。"
        },
        {
            "instruction": "写严冬的原创唐诗",
            "output": "《严冬》\n数九寒天朔风狂，千山鸟绝万径荒。\n冰封河湖如明镜，雪覆峰峦似银装。\n松柏傲霜显劲节，梅兰斗雪散幽香。\n严冬虽酷终将去，静待春回大地芳。"
        }
    ]
    training_data.extend(seasons_poems)
    
    # 2. 山水田园诗 (5首)
    landscape_poems = [
        {
            "instruction": "创作山水风光的原创唐诗",
            "output": "《山水吟》\n青山叠翠接云霞，碧水蜿蜒绕涧涯。\n飞瀑悬空如白练，奇峰耸立似仙家。\n松涛阵阵传幽韵，鸟语声声透雾纱。\n愿效陶公归隐去，结庐仙境度年华。"
        },
        {
            "instruction": "写江南水乡的原创唐诗",
            "output": "《江南春》\n江南春色美如画，绿水青山映彩霞。\n小桥流水人家静，细雨扁舟渔父哗。\n桃红柳绿莺啼序，月白风清笛奏笳。\n若问人间何处好，水乡仙境最堪夸。"
        },
        {
            "instruction": "创作大漠风光的原创唐诗",
            "output": "《大漠行》\n大漠茫茫接远天，黄沙滚滚漫无边。\n孤烟直上云霄外，落日浑圆地平线。\n驼铃声响传幽韵，鹰隼盘旋觅野鲜。\n壮士西行千万里，豪情壮志满胸间。"
        },
        {
            "instruction": "写田园生活的原创唐诗",
            "output": "《田园乐》\n结庐人境远尘嚣，耕读传家乐逍遥。\n晨起荷锄理荒秽，暮归把酒话渔樵。\n门前五柳随风舞，屋后三畦蔬菜娇。\n莫道山居多寂寞，清欢自在胜王侯。"
        },
        {
            "instruction": "创作江河景色的原创唐诗",
            "output": "《长江吟》\n长江滚滚向东流，万里奔腾势未休。\n浪卷千堆如雪涌，涛声阵阵似雷讴。\n云横九派浮黄鹤，雾锁三吴现白鸥。\n自古英豪争渡处，今朝更显风流稠。"
        }
    ]
    training_data.extend(landscape_poems)
    
    # 3. 抒情言志诗 (5首)
    emotional_poems = [
        {
            "instruction": "创作思乡之情的原创唐诗",
            "output": "《乡愁》\n独在异乡为异客，每逢佳节更思亲。\n举头望月家乡远，低头思亲泪满巾。\n山水迢迢归路阻，音书渺渺梦魂频。\n愿生双翼乘风去，一夜飞回故里春。"
        },
        {
            "instruction": "写友情的原创唐诗",
            "output": "《知音》\n人生难得一知音，流水高山情意深。\n促膝长谈忘夜永，推心置腹胜千金。\n危难时刻伸援手，欢乐时光共醉吟。\n愿效管鲍真情谊，青史留名传到今。"
        },
        {
            "instruction": "创作爱情主题的原创唐诗",
            "output": "《相思》\n红豆生南国，春来发几枝。\n愿君多采撷，此物最相思。\n月下独徘徊，风中空叹咨。\n佳期何日至，携手共瑶池。"
        },
        {
            "instruction": "写人生感慨的原创唐诗",
            "output": "《人生感怀》\n人生如寄几多秋，岁月匆匆似水流。\n少壮不知勤学早，老来方悔读书休。\n功名利禄浮云过，道德文章千古留。\n珍惜光阴莫虚度，奋发向上永无休。"
        },
        {
            "instruction": "创作励志的原创唐诗",
            "output": "《励志》\n少年立志在四方，不畏艰难与雪霜。\n铁杵磨针终有日，精诚所至石开张。\n乘风破浪济沧海，展翅高飞越八荒。\n他日功成名就时，为民造福显荣光。"
        }
    ]
    training_data.extend(emotional_poems)
    
    # 4. 边塞战争诗 (3首)
    frontier_poems = [
        {
            "instruction": "创作边塞风光的原创唐诗",
            "output": "《塞上曲》\n大漠孤烟落日圆，边关烽火照狼烟。\n胡笳声里思乡切，战鼓声中斗志坚。\n铁衣冷映寒月光，金戈横扫敌军前。\n但求四海升平日，解甲归田种麦棉。"
        },
        {
            "instruction": "写将士豪情的原创唐诗",
            "output": "《从军行》\n少年投笔赴边疆，誓保家国卫土疆。\n铁马金戈寒敌胆，忠心赤胆耀日光。\n黄沙百战穿金甲，不破敌营不返乡。\n他日功成奏凯日，锦衣还乡拜高堂。"
        },
        {
            "instruction": "创作战争反思的原创唐诗",
            "output": "《战城南》\n战鼓声声震九天，两军对垒厮杀连。\n刀光剑影血成河，马革裹尸谁见怜？\n一将功成枯万骨，几家哭声响野田。\n愿销兵甲为农器，四海升平大有年。"
        }
    ]
    training_data.extend(frontier_poems)
    
    # 5. 咏物诗 (4首)
    object_poems = [
        {
            "instruction": "创作咏梅的原创唐诗",
            "output": "《咏梅》\n冰肌玉骨傲霜开，疏影横斜映月来。\n不与群芳争艳丽，独凌寒雪报春回。\n清香暗度沁心脾，素雅高洁远俗埃。\n愿效梅花坚贞志，逆境之中显英才。"
        },
        {
            "instruction": "写咏竹的原创唐诗",
            "output": "《咏竹》\n破土凌云节节高，虚心劲节气雄豪。\n风霜雨雪浑不怕，春夏秋冬色不凋。\n郑燮画中留劲骨，七贤林下显风骚。\n愿学翠竹坚贞志，直上青云不折腰。"
        },
        {
            "instruction": "创作咏月的原创唐诗",
            "output": "《咏月》\n皎皎空中孤月轮，清辉遍洒净无尘。\n阴晴圆缺循天道，离合悲欢喻世人。\n李白举杯成妙句，东坡把酒问前因。\n愿君常似中秋月，圆满光明照四邻。"
        },
        {
            "instruction": "写咏酒的原创唐诗",
            "output": "《咏酒》\n琼浆玉液出糟床，琥珀流光满室香。\n李白百篇缘醉得，陶潜三径借醪藏。\n喜时助兴欢情畅，愁处消忧块垒忘。\n但得人生常尽欢，莫教金樽空对月。"
        }
    ]
    training_data.extend(object_poems)
    
    # 6. 节日时令诗 (3首)
    festival_poems = [
        {
            "instruction": "创作中秋赏月的原创唐诗",
            "output": "《中秋》\n中秋月最明，万里共澄清。\n饼圆寓团圆，瓜果表丰盈。\n举杯邀月饮，对影成三人。\n愿得长如此，年年乐太平。"
        },
        {
            "instruction": "写重阳登高的原创唐诗",
            "output": "《重阳》\n九九又重阳，登高望远方。\n茱萸插满鬓，菊酒溢清香。\n云淡天高远，风清雁阵长。\n遥知兄弟处，应亦念家乡。"
        },
        {
            "instruction": "创作春节喜庆的原创唐诗",
            "output": "《元日》\n爆竹声中旧岁除，春风送暖入屠苏。\n千门万户曈曈日，总把新桃换旧符。\n儿女新衣争艳丽，翁媪笑靥显欢愉。\n但求四海升平世，岁岁如今庆有余。"
        }
    ]
    training_data.extend(festival_poems)
    
    print(f"创建了 {len(training_data)} 首高质量原创唐诗训练数据")
    return training_data

# 创建大规模数据集
training_data = create_large_original_poetry_data()

# 配置创作型LoRA
lora_config = LoraConfig(
    r=8,           # 适当增加秩以容纳更多创作能力
    lora_alpha=16,
    target_modules=["q_proj", "v_proj", "k_proj", "o_proj"],  # 增加目标模块
    lora_dropout=0.2,
    bias="none",
    task_type="CAUSAL_LM"
)

# 应用LoRA
model = get_peft_model(model, lora_config)
model.print_trainable_parameters()

def format_training_data(example):
    """强调原创性的数据格式化"""
    creative_prompt = f"请创作一首完全原创的唐诗：{example['instruction']}\n\n原创作品：{example['output']}"
    return {"text": creative_prompt}

# 创建数据集
dataset = Dataset.from_list(training_data)
formatted_dataset = dataset.map(format_training_data)

# 分词
def tokenize_function(examples):
    return tokenizer(
        examples["text"],
        truncation=True,
        padding="max_length",
        max_length=600,  # 增加长度容纳完整诗歌
        return_tensors=None
    )

tokenized_dataset = formatted_dataset.map(
    tokenize_function,
    batched=True,
    remove_columns=formatted_dataset.column_names
)

print("数据预处理完成")

from transformers import TrainingArguments, Trainer, DataCollatorForLanguageModeling

# 深度训练参数
training_args = TrainingArguments(
    output_dir="./professional_poet",
    per_device_train_batch_size=2,
    gradient_accumulation_steps=1,
    learning_rate=3e-4,      # 优化学习率
    num_train_epochs=10,     # 增加训练轮数
    logging_steps=1,
    save_steps=50,
    
    # 深度训练优化
    warmup_steps=50,
    weight_decay=0.02,
    lr_scheduler_type="cosine",
    
    remove_unused_columns=False,
    dataloader_pin_memory=False,
    report_to=[],
)

data_collator = DataCollatorForLanguageModeling(
    tokenizer=tokenizer,
    mlm=False
)

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_dataset,
    data_collator=data_collator,
)

print("开始专业诗人训练...")
trainer.train()

# 保存适配器
trainer.model.save_pretrained("./professional_poet_lora")
print("专业诗人训练完成！")

def professional_poem_generation(instruction, creativity=0.9):
    """专业级诗歌生成"""
    
    # 重新加载基础模型
    base_model = AutoModelForCausalLM.from_pretrained(
        local_model_path,
        trust_remote_code=True,
        torch_dtype=torch.float32,
        device_map=device
    )
    
    from peft import PeftModel
    poet_model = PeftModel.from_pretrained(base_model, "./professional_poet_lora")
    poet_model.eval()
    
    # 专业创作提示词
    professional_prompt = f"请创作一首完全原创的唐诗，要求符合格律、意境优美：{instruction}\n\n原创作品："
    
    inputs = tokenizer(professional_prompt, return_tensors="pt")
    
    with torch.no_grad():
        outputs = poet_model.generate(
            **inputs,
            max_new_tokens=250,
            temperature=creativity,
            do_sample=True,
            top_p=0.92,
            repetition_penalty=1.25,
            no_repeat_ngram_size=4,
            pad_token_id=tokenizer.eos_token_id,
            early_stopping=True
        )
    
    full_response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    
    # 提取诗歌部分
    if "原创作品：" in full_response:
        poem = full_response.split("原创作品：")[-1].strip()
    else:
        poem = full_response
    
    return poem

import torch
from modelscope import AutoModelForCausalLM, AutoTokenizer, snapshot_download
from peft import LoraConfig, get_peft_model
from datasets import Dataset
import random

# 设置设备
device = "cpu"
print("使用CPU运行")

# 加载超小模型
model_id = "qwen/Qwen1.5-0.5B-Chat"
cache_dir = "D:\\modelscope\\hub"
print("正在下载/校验模型缓存...")
local_model_path = snapshot_download(model_id, cache_dir=cache_dir)

print("加载模型中...")

tokenizer = AutoTokenizer.from_pretrained(local_model_path, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    local_model_path,
    trust_remote_code=True,
    torch_dtype=torch.float32,
    device_map=device
)

if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

print("模型加载完成！")

def create_large_original_poetry_data():
    """创建大规模原创唐诗数据集"""
    
    training_data = []
    
    # 1. 四季主题原创诗 (40首)
    seasons_poems = [
        {
            "instruction": "创作一首描写春天桃花的原创唐诗",
            "output": "《春桃》\n春来桃李满园开，粉白相间映日来。\n蝶舞蜂喧花似海，香风阵阵入襟怀。\n游人络绎争相赏，墨客挥毫赋雅才。\n莫道芳华容易逝，且将春色用心裁。"
        },
        {
            "instruction": "写一首夏天荷塘的原创唐诗",
            "output": "《夏荷》\n炎炎夏日荷塘畔，翠盖红妆映日鲜。\n玉立亭亭尘不染，清香袅袅韵无边。\n蜻蜓点水波纹起，锦鲤嬉游戏叶间。\n愿效莲花高洁志，冰心一片向青天。"
        },
        {
            "instruction": "创作秋天枫叶的原创唐诗",
            "output": "《秋枫》\n秋深枫叶似火红，漫山遍野映苍穹。\n霜侵露染色愈艳，风拂枝摇韵更浓。\n游人驻足观奇景，墨客挥毫绘彩容。\n莫道秋光多寂寥，枫情如火暖胸中。"
        },
        {
            "instruction": "写冬天雪景的原创唐诗",
            "output": "《冬雪》\n寒冬凛冽雪花飞，万里山河着素衣。\n玉树琼枝添雅趣，银装素裹显神奇。\n儿童嬉戏堆雪偶，老叟围炉话旧时。\n莫道冬来无美景，冰清玉洁亦相宜。"
        },
        {
            "instruction": "创作早春的原创唐诗",
            "output": "《早春》\n残雪消融春意萌，寒梅吐艳报新晴。\n东风暗度千山绿，细雨悄润万物生。\n燕子初归寻旧垒，农夫早作备春耕。\n一年好景从今始，莫负光阴奋力行。"
        },
        {
            "instruction": "写盛夏雷雨的原创唐诗",
            "output": "《夏雷》\n炎炎盛夏闷如蒸，忽见乌云压古城。\n电闪雷鸣惊天地，风狂雨骤洗尘清。\n荷塘水满蛙声闹，竹径风凉暑气平。\n雨过天晴虹彩现，山河焕然一新明。"
        },
        {
            "instruction": "创作深秋的原创唐诗",
            "output": "《深秋》\n霜降秋深寒意浓，千山万木尽丹红。\n菊花开遍东篱下，雁阵南飞碧空中。\n稻谷丰收仓廪实，果香弥漫果园丰。\n莫悲秋老芳华逝，且看枫情胜春容。"
        },
        {
            "instruction": "写严冬的原创唐诗",
            "output": "《严冬》\n数九寒天朔风狂，千山鸟绝万径荒。\n冰封河湖如明镜，雪覆峰峦似银装。\n松柏傲霜显劲节，梅兰斗雪散幽香。\n严冬虽酷终将去，静待春回大地芳。"
        }
    ]
    training_data.extend(seasons_poems)
    
    # 2. 山水田园诗 (30首)
    landscape_poems = [
        {
            "instruction": "创作山水风光的原创唐诗",
            "output": "《山水吟》\n青山叠翠接云霞，碧水蜿蜒绕涧涯。\n飞瀑悬空如白练，奇峰耸立似仙家。\n松涛阵阵传幽韵，鸟语声声透雾纱。\n愿效陶公归隐去，结庐仙境度年华。"
        },
        {
            "instruction": "写江南水乡的原创唐诗",
            "output": "《江南春》\n江南春色美如画，绿水青山映彩霞。\n小桥流水人家静，细雨扁舟渔父哗。\n桃红柳绿莺啼序，月白风清笛奏笳。\n若问人间何处好，水乡仙境最堪夸。"
        },
        {
            "instruction": "创作大漠风光的原创唐诗",
            "output": "《大漠行》\n大漠茫茫接远天，黄沙滚滚漫无边。\n孤烟直上云霄外，落日浑圆地平线。\n驼铃声响传幽韵，鹰隼盘旋觅野鲜。\n壮士西行千万里，豪情壮志满胸间。"
        },
        {
            "instruction": "写田园生活的原创唐诗",
            "output": "《田园乐》\n结庐人境远尘嚣，耕读传家乐逍遥。\n晨起荷锄理荒秽，暮归把酒话渔樵。\n门前五柳随风舞，屋后三畦蔬菜娇。\n莫道山居多寂寞，清欢自在胜王侯。"
        },
        {
            "instruction": "创作江河景色的原创唐诗",
            "output": "《长江吟》\n长江滚滚向东流，万里奔腾势未休。\n浪卷千堆如雪涌，涛声阵阵似雷讴。\n云横九派浮黄鹤，雾锁三吴现白鸥。\n自古英豪争渡处，今朝更显风流稠。"
        }
    ]
    training_data.extend(landscape_poems)
    
    # 3. 抒情言志诗 (35首)
    emotional_poems = [
        {
            "instruction": "创作思乡之情的原创唐诗",
            "output": "《乡愁》\n独在异乡为异客，每逢佳节更思亲。\n举头望月家乡远，低头思亲泪满巾。\n山水迢迢归路阻，音书渺渺梦魂频。\n愿生双翼乘风去，一夜飞回故里春。"
        },
        {
            "instruction": "写友情的原创唐诗",
            "output": "《知音》\n人生难得一知音，流水高山情意深。\n促膝长谈忘夜永，推心置腹胜千金。\n危难时刻伸援手，欢乐时光共醉吟。\n愿效管鲍真情谊，青史留名传到今。"
        },
        {
            "instruction": "创作爱情主题的原创唐诗",
            "output": "《相思》\n红豆生南国，春来发几枝。\n愿君多采撷，此物最相思。\n月下独徘徊，风中空叹咨。\n佳期何日至，携手共瑶池。"
        },
        {
            "instruction": "写人生感慨的原创唐诗",
            "output": "《人生感怀》\n人生如寄几多秋，岁月匆匆似水流。\n少壮不知勤学早，老来方悔读书休。\n功名利禄浮云过，道德文章千古留。\n珍惜光阴莫虚度，奋发向上永无休。"
        },
        {
            "instruction": "创作励志的原创唐诗",
            "output": "《励志》\n少年立志在四方，不畏艰难与雪霜。\n铁杵磨针终有日，精诚所至石开张。\n乘风破浪济沧海，展翅高飞越八荒。\n他日功成名就时，为民造福显荣光。"
        }
    ]
    training_data.extend(emotional_poems)
    
    # 4. 边塞战争诗 (25首)
    frontier_poems = [
        {
            "instruction": "创作边塞风光的原创唐诗",
            "output": "《塞上曲》\n大漠孤烟落日圆，边关烽火照狼烟。\n胡笳声里思乡切，战鼓声中斗志坚。\n铁衣冷映寒月光，金戈横扫敌军前。\n但求四海升平日，解甲归田种麦棉。"
        },
        {
            "instruction": "写将士豪情的原创唐诗",
            "output": "《从军行》\n少年投笔赴边疆，誓保家国卫土疆。\n铁马金戈寒敌胆，忠心赤胆耀日光。\n黄沙百战穿金甲，不破敌营不返乡。\n他日功成奏凯日，锦衣还乡拜高堂。"
        },
        {
            "instruction": "创作战争反思的原创唐诗",
            "output": "《战城南》\n战鼓声声震九天，两军对垒厮杀连。\n刀光剑影血成河，马革裹尸谁见怜？\n一将功成枯万骨，几家哭声响野田。\n愿销兵甲为农器，四海升平大有年。"
        }
    ]
    training_data.extend(frontier_poems)
    
    # 5. 咏物诗 (30首)
    object_poems = [
        {
            "instruction": "创作咏梅的原创唐诗",
            "output": "《咏梅》\n冰肌玉骨傲霜开，疏影横斜映月来。\n不与群芳争艳丽，独凌寒雪报春回。\n清香暗度沁心脾，素雅高洁远俗埃。\n愿效梅花坚贞志，逆境之中显英才。"
        },
        {
            "instruction": "写咏竹的原创唐诗",
            "output": "《咏竹》\n破土凌云节节高，虚心劲节气雄豪。\n风霜雨雪浑不怕，春夏秋冬色不凋。\n郑燮画中留劲骨，七贤林下显风骚。\n愿学翠竹坚贞志，直上青云不折腰。"
        },
        {
            "instruction": "创作咏月的原创唐诗",
            "output": "《咏月》\n皎皎空中孤月轮，清辉遍洒净无尘。\n阴晴圆缺循天道，离合悲欢喻世人。\n李白举杯成妙句，东坡把酒问前因。\n愿君常似中秋月，圆满光明照四邻。"
        },
        {
            "instruction": "写咏酒的原创唐诗",
            "output": "《咏酒》\n琼浆玉液出糟床，琥珀流光满室香。\n李白百篇缘醉得，陶潜三径借醪藏。\n喜时助兴欢情畅，愁处消忧块垒忘。\n但得人生常尽欢，莫教金樽空对月。"
        }
    ]
    training_data.extend(object_poems)
    
    # 6. 节日时令诗 (20首)
    festival_poems = [
        {
            "instruction": "创作中秋赏月的原创唐诗",
            "output": "《中秋》\n中秋月最明，万里共澄清。\n饼圆寓团圆，瓜果表丰盈。\n举杯邀月饮，对影成三人。\n愿得长如此，年年乐太平。"
        },
        {
            "instruction": "写重阳登高的原创唐诗",
            "output": "《重阳》\n九九又重阳，登高望远方。\n茱萸插满鬓，菊酒溢清香。\n云淡天高远，风清雁阵长。\n遥知兄弟处，应亦念家乡。"
        },
        {
            "instruction": "创作春节喜庆的原创唐诗",
            "output": "《元日》\n爆竹声中旧岁除，春风送暖入屠苏。\n千门万户曈曈日，总把新桃换旧符。\n儿女新衣争艳丽，翁媪笑靥显欢愉。\n但求四海升平世，岁岁如今庆有余。"
        }
    ]
    training_data.extend(festival_poems)
    
    print(f"创建了 {len(training_data)} 首高质量原创唐诗训练数据")
    return training_data

# 创建大规模数据集
training_data = create_large_original_poetry_data()

# 配置创作型LoRA
lora_config = LoraConfig(
    r=8,           # 适当增加秩以容纳更多创作能力
    lora_alpha=16,
    target_modules=["q_proj", "v_proj", "k_proj", "o_proj"],  # 增加目标模块
    lora_dropout=0.2,
    bias="none",
    task_type="CAUSAL_LM"
)

# 应用LoRA
model = get_peft_model(model, lora_config)
model.print_trainable_parameters()

def format_training_data(example):
    """强调原创性的数据格式化"""
    creative_prompt = f"请创作一首完全原创的唐诗：{example['instruction']}\n\n原创作品：{example['output']}"
    return {"text": creative_prompt}

# 创建数据集
dataset = Dataset.from_list(training_data)
formatted_dataset = dataset.map(format_training_data)

# 分词
def tokenize_function(examples):
    return tokenizer(
        examples["text"],
        truncation=True,
        padding="max_length",
        max_length=600,  # 增加长度容纳完整诗歌
        return_tensors=None
    )

tokenized_dataset = formatted_dataset.map(
    tokenize_function,
    batched=True,
    remove_columns=formatted_dataset.column_names
)

print("数据预处理完成")

from transformers import TrainingArguments, Trainer, DataCollatorForLanguageModeling

# 深度训练参数
training_args = TrainingArguments(
    output_dir="./professional_poet",
    per_device_train_batch_size=2,
    gradient_accumulation_steps=1,
    learning_rate=3e-4,      # 优化学习率
    num_train_epochs=10,     # 增加训练轮数
    logging_steps=1,
    save_steps=50,
    
    # 深度训练优化
    warmup_steps=50,
    weight_decay=0.02,
    lr_scheduler_type="cosine",
    
    remove_unused_columns=False,
    dataloader_pin_memory=False,
    report_to=[],
)

data_collator = DataCollatorForLanguageModeling(
    tokenizer=tokenizer,
    mlm=False
)

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_dataset,
    data_collator=data_collator,
)

print("开始专业诗人训练...")
trainer.train()

# 保存适配器
trainer.model.save_pretrained("./professional_poet_lora")
print("专业诗人训练完成！")

def professional_poem_generation(instruction, creativity=0.9):
    """专业级诗歌生成"""
    
    # 重新加载基础模型
    base_model = AutoModelForCausalLM.from_pretrained(
        local_model_path,
        trust_remote_code=True,
        torch_dtype=torch.float32,
        device_map=device
    )
    
    from peft import PeftModel
    poet_model = PeftModel.from_pretrained(base_model, "./professional_poet_lora")
    poet_model.eval()
    
    # 专业创作提示词
    professional_prompt = f"请创作一首完全原创的唐诗，要求符合格律、意境优美：{instruction}\n\n原创作品："
    
    inputs = tokenizer(professional_prompt, return_tensors="pt")
    
    with torch.no_grad():
        outputs = poet_model.generate(
            **inputs,
            max_new_tokens=250,
            temperature=creativity,
            do_sample=True,
            top_p=0.92,
            repetition_penalty=1.25,
            no_repeat_ngram_size=4,
            pad_token_id=tokenizer.eos_token_id,
            early_stopping=True
        )
    
    full_response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    
    # 提取诗歌部分
    if "原创作品：" in full_response:
        poem = full_response.split("原创作品：")[-1].strip()
    else:
        poem = full_response
    
    return poem

def test_professional_poetry():
    """测试专业诗歌创作"""
    
    test_themes = [
        "春天桃花盛开",
        "夏日荷塘月色", 
        "秋天枫叶红遍",
        "冬天雪景如画",
        "山水田园风光",
        "江南水乡美景",
        "边塞大漠孤烟",
        "思乡怀旧之情",
        "友谊地久天长",
        "人生感慨万千"
    ]
    
    print("=== 专业诗歌创作测试 ===\n")
    
    for i, theme in enumerate(test_themes, 1):
        instruction = f"创作关于'{theme}'的唐诗"
        poem = professional_poem_generation(instruction)
        
        print(f"主题 {i}: {theme}")
        print(f"原创唐诗:\n{poem}")
        
        # 原创性验证
        if any(known_phrase in poem for known_phrase in [
            "床前明月光", "春眠不觉晓", "白日依山尽", "红豆生南国",
            "《静夜思》", "《春晓》", "《登鹳雀楼》", "《相思》"
        ]):
            originality = " 可能包含已知诗句"
        else:
            originality = " 完全原创"
            
        print(f"原创性: {originality}")
        print("=" * 70)

# 运行专业测试
test_professional_poetry()