Add network admin LLM training script

network_admin_llm_train.py

@@ -0,0 +1,333 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+'''
+Network Admin LLM - QLoRA Fine-tuning Script
+=============================================
+Base Model: microsoft/Phi-4-mini-instruct
+Method: QLoRA SFT (4-bit quantization + LoRA)
+Datasets: NetEval + Telecom Intent Config
+
+Run locally with GPU:
+    pip install transformers trl peft bitsandbytes accelerate datasets trackio
+    python network_admin_llm_train.py
+
+Or on Google Colab:
+    !pip install transformers trl peft bitsandbytes accelerate datasets trackio
+    %cd /content
+    !python network_admin_llm_train.py
+
+Author: Network Admin LLM Project
+'''
+
+import os
+import sys
+import torch
+from datetime import datetime
+
+# ============== CONFIGURATION ==============
+# 請修改以下設定
+MODEL_NAME = 'microsoft/Phi-4-mini-instruct'
+HF_USERNAME = 'YOUR_HF_USERNAME'  # 改成你的 HuggingFace 用戶名
+HF_TOKEN = os.environ.get('HF_TOKEN', 'YOUR_HF_TOKEN')  # HF token for upload
+
+# 訓練超參數
+TRAINING_CONFIG = {
+    'learning_rate': 2e-4,          # LoRA 需要較高學習率
+    'num_epochs': 3,
+    'batch_size': 4,
+    'gradient_accumulation': 4,      # effective batch = 16
+    'max_seq_length': 2048,
+    'lora_r': 16,
+    'lora_alpha': 32,
+    'lora_dropout': 0.05,
+    'warmup_ratio': 0.1,
+}
+
+OUTPUT_DIR = f'{HF_USERNAME}/network-admin-phi4-mini'
+# ===========================================
+
+def print_section(title):
+    print(f'\n{"="*60}')
+    print(f' {title}')
+    print('='*60)
+
+def install_dependencies():
+    '''檢查並安裝依賴'''
+    print_section('CHECKING DEPENDENCIES')
+    
+    required = ['transformers', 'trl', 'peft', 'bitsandbytes', 'accelerate', 'datasets', 'trackio']
+    missing = []
+    
+    for pkg in required:
+        try:
+            __import__(pkg.replace('-', '_'))
+            print(f'✅ {pkg}')
+        except ImportError:
+            missing.append(pkg)
+            print(f'❌ {pkg} - 需要安裝')
+    
+    if missing:
+        print(f'\n請運行: pip install {" ".join(missing)}')
+        return False
+    return True
+
+def load_and_prepare_datasets():
+    '''載入並轉換數據集'''
+    from datasets import load_dataset, concatenate_datasets
+    
+    print_section('LOADING DATASETS')
+    
+    # 1. 載入 NetEval 考試題庫
+    print('📚 載入 NetEval 考試題庫...')
+    neteval_dataset = load_dataset('NASP/neteval-exam', split='train')
+    print(f'   NetEval: {len(neteval_dataset)} 題')
+    
+    def convert_neteval(example):
+        '''將 Q&A 格式轉換為對話格式'''
+        question = example['Question']
+        options = f'\nA. {example.get("A", "")}\nB. {example.get("B", "")}\nC. {example.get("C", "")}\nD. {example.get("D", "")}'
+        
+        answer = f'正確答案是: {example["Answer"]}'
+        if example.get('Explanation'):
+            answer += f'\n\n📖 解說: {example["Explanation"]}'
+        
+        return {
+            'messages': [
+                {'role': 'system', 'content': '你是一位網路管理專家。請回答關於網路、安全、路由、交換機、VLAN、防火牆等IT基礎設施的問題。'},
+                {'role': 'user', 'content': f'{question}{options}'},
+                {'role': 'assistant', 'content': answer}
+            ]
+        }
+    
+    neteval_converted = neteval_dataset.map(
+        convert_neteval,
+        remove_columns=neteval_dataset.column_names,
+        desc='轉換 NetEval 格式'
+    )
+    
+    # 2. 載入電信意圖配置數據集
+    print('📚 載入電信意圖配置數據集...')
+    telecom_dataset = load_dataset('nraptisss/telecom-intent-config-sft-10k', split='train')
+    print(f'   Telecom: {len(telecom_dataset)} 條')
+    
+    telecom_messages = telecom_dataset.map(
+        lambda x: {'messages': x['messages']},
+        remove_columns=[c for c in telecom_dataset.column_names if c != 'messages']
+    )
+    
+    # 3. 合併數據集
+    print('🔄 合併數據集...')
+    combined = concatenate_datasets([neteval_converted, telecom_messages])
+    split_data = combined.train_test_split(test_size=0.1, seed=42)
+    
+    train_ds = split_data['train']
+    eval_ds = split_data['test']
+    
+    print(f'\n📊 數據集統計:')
+    print(f'   訓練集: {len(train_ds)} 條')
+    print(f'   驗證集: {len(eval_ds)} 條')
+    print(f'   總計: {len(combined)} 條')
+    
+    return train_ds, eval_ds
+
+def setup_model_and_tokenizer():
+    '''設置模型和 tokenizer'''
+    from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
+    from peft import LoraConfig, prepare_model_for_kbit_training, get_peft_model
+    
+    print_section('LOADING MODEL')
+    print(f'🤖 模型: {MODEL_NAME}')
+    
+    # Tokenizer
+    print('\n📝 載入 Tokenizer...')
+    tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
+    tokenizer.pad_token = tokenizer.eos_token
+    tokenizer.padding_side = 'right'
+    print(f'   Vocab size: {len(tokenizer):,}')
+    
+    # QLoRA 配置 (4-bit)
+    print('\n��� 配置 QLoRA (4-bit)...')
+    bnb_config = BitsAndBytesConfig(
+        load_in_4bit=True,
+        bnb_4bit_quant_type='nf4',          # Normalized Float4
+        bnb_4bit_compute_dtype=torch.bfloat16,
+        bnb_4bit_use_double_quant=True,      # 嵌套量化
+    )
+    
+    # 載入模型
+    print('📥 載入模型 (4-bit)...')
+    model = AutoModelForCausalLM.from_pretrained(
+        MODEL_NAME,
+        quantization_config=bnb_config,
+        device_map='auto',
+        trust_remote_code=True,
+    )
+    
+    # 準備 kbit 訓練
+    model = prepare_model_for_kbit_training(model)
+    print('✅ 模型準備完成')
+    
+    # LoRA 配置
+    print('\n🔧 配置 LoRA...')
+    lora_config = LoraConfig(
+        r=TRAINING_CONFIG['lora_r'],
+        lora_alpha=TRAINING_CONFIG['lora_alpha'],
+        lora_dropout=TRAINING_CONFIG['lora_dropout'],
+        bias='none',
+        task_type='CAUSAL_LM',
+        target_modules=[
+            'q_proj', 'k_proj', 'v_proj', 'o_proj',  # Attention
+            'gate_proj', 'up_proj', 'down_proj',      # MLP
+        ],
+        modules_to_save=['lm_head', 'embed_tokens'],
+    )
+    
+    # 應用 LoRA
+    model = get_peft_model(model, lora_config)
+    model.print_trainable_parameters()
+    
+    return model, tokenizer, lora_config
+
+def setup_trainer(model, tokenizer, train_ds, eval_ds, lora_config):
+    '''設置訓練器'''
+    from trl import SFTTrainer, SFTConfig
+    
+    print_section('CONFIGURING TRAINER')
+    
+    # 生成運行名稱
+    run_name = f'phi4-netadmin-{datetime.now().strftime("%m%d-%H%M")}'
+    
+    # 嘗試初始化 trackio
+    try:
+        import trackio
+        trackio.init(project='network-admin-llm', experiment='qlora-sft', run_name=run_name)
+        print('✅ Trackio 初始化成功')
+        report_to = ['trackio']
+    except Exception as e:
+        print(f'⚠️ Trackio 初始化失敗: {e}')
+        report_to = ['none']
+    
+    # SFT 配置
+    training_args = SFTConfig(
+        # 學習率
+        learning_rate=TRAINING_CONFIG['learning_rate'],
+        lr_scheduler_type='cosine',
+        warmup_ratio=TRAINING_CONFIG['warmup_ratio'],
+        
+        # 訓練
+        num_train_epochs=TRAINING_CONFIG['num_epochs'],
+        per_device_train_batch_size=TRAINING_CONFIG['batch_size'],
+        gradient_accumulation_steps=TRAINING_CONFIG['gradient_accumulation'],
+        max_seq_length=TRAINING_CONFIG['max_seq_length'],
+        
+        # 記憶體優化
+        gradient_checkpointing=True,
+        bf16=True,
+        fp16=False,
+        
+        # 輸出
+        output_dir='./output',
+        logging_steps=10,
+        save_steps=500,
+        save_total_limit=2,
+        evaluation_strategy='steps',
+        eval_steps=500,
+        
+        # Hub 上傳
+        push_to_hub=True,
+        hub_model_id=OUTPUT_DIR,
+        hub_strategy='checkpoint',
+        
+        # 監控
+        report_to=report_to,
+        logging_strategy='steps',
+        logging_first_step=True,
+        
+        # 雜項
+        remove_unused_columns=False,
+        dataloader_num_workers=4,
+        seed=42,
+    )
+    
+    # 創建 trainer
+    trainer = SFTTrainer(
+        model=model,
+        args=training_args,
+        train_dataset=train_ds,
+        eval_dataset=eval_ds,
+        processing_class=tokenizer,
+        peft_config=lora_config,
+    )
+    
+    return trainer, run_name
+
+def train_model(trainer):
+    '''執行訓練'''
+    print_section('STARTING TRAINING')
+    print('🚀 開始訓練...')
+    print('   (按 Ctrl+C 可隨時中斷)')
+    print()
+    
+    try:
+        trainer.train()
+        print('\n✅ 訓練完成!')
+        return True
+    except KeyboardInterrupt:
+        print('\n⚠️ 訓練被用戶中斷')
+        return False
+    except Exception as e:
+        print(f'\n❌ 訓練失敗: {e}')
+        raise
+
+def save_and_upload(trainer):
+    '''保存並上傳模型'''
+    print_section('SAVING & UPLOADING')
+    
+    try:
+        print('📤 上傳模型到 HuggingFace Hub...')
+        trainer.push_to_hub()
+        print(f'\n✅ 模型已上傳!')
+        print(f'🔗 連結: https://huggingface.co/{OUTPUT_DIR}')
+    except Exception as e:
+        print(f'\n⚠️ 上傳失敗: {e}')
+        print('模型已保存在 ./output 目錄')
+
+def main():
+    '''主函數'''
+    print('''
+    ╔═══════════════════════════════════════════════════════════╗
+    ║       Network Admin LLM - QLoRA Fine-tuning              ║
+    ║       Base: microsoft/Phi-4-mini-instruct                ║
+    ╚═══════════════════════════════════════════════════════════╝
+    ''')
+    
+    # 檢查 GPU
+    print(f'🖥️  GPU: {torch.cuda.get_device_name(0) if torch.cuda.is_available() else "無 GPU"}')
+    if torch.cuda.is_available():
+        print(f'   Memory: {torch.cuda.get_device_properties(0).total_memory / 1024**3:.1f} GB')
+    
+    # 安裝依賴
+    if not install_dependencies():
+        sys.exit(1)
+    
+    # 載入數據
+    train_ds, eval_ds = load_and_prepare_datasets()
+    
+    # 設置模型
+    model, tokenizer, lora_config = setup_model_and_tokenizer()
+    
+    # 設置 trainer
+    trainer, run_name = setup_trainer(model, tokenizer, train_ds, eval_ds, lora_config)
+    
+    # 訓練
+    success = train_model(trainer)
+    
+    # 保存
+    if success:
+        save_and_upload(trainer)
+    
+    print_section('DONE')
+    print(f'Run name: {run_name}')
+
+if __name__ == '__main__':
+    main()