LLMAlignment/Hello.py at master · AgentZero2002/LLMAlignment · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
"""
LLM 从基础到前沿 — Part III: 对齐与后训练

80 章 = 5 Parts × 16 章:
    Part I:   基础与架构 (01-16)     ✅ LLMBasic
    Part II:  训练与推理 (17-32)     ✅ LLMTrainingAndInference
    Part III: 对齐与后训练 (33-48)    ← 当前 LLMAlignment
    Part IV:  高级能力 (49-64)        LLMAdvance
    Part V:   前沿与应用 (65-80)    LLMFrontier

快速运行:
    python3 Hello.py
    python3 33_sft/sft.py
    python3 38_dpo/dpo.py
    python3 43_peft_lora/peft_lora.py
"""

def main():
    print("╔══════════════════════════════════════════════════════╗")
    print("║   LLM 从基础到前沿 — 对齐与后训练 (Part III)       ║")
    print("║   清华大学 · 计算机科学与技术系                      ║")
    print("╚══════════════════════════════════════════════════════╝\n")

    chapters = [
        ("33", "SFT — 监督微调",           "从续写器到助手 · Loss Masking"),
        ("34", "SFT 数据工程",              "收集·清洗·合成·Self-Instruct"),
        ("35", "RLHF 概览",                "三阶段流水线 · 偏好 > 评分"),
        ("36", "奖励模型",                 "Bradley-Terry · 奖励 Hacking"),
        ("37", "PPO — 近端策略优化",       "KL约束 · Clip · GAE"),
        ("38", "DPO — 直接偏好优化",       "闭式解 · 跳过RM+PPO"),
        ("39", "DPO 变体全家桶",           "IPO · KTO · ORPO · SimPO"),
        ("40", "Constitutional AI & RLAIF","AI宪法 · AI替代人类"),
        ("41", "安全与红队测试",           "攻击·越狱·安全训练"),
        ("42", "LLM 评估体系",             "MMLU · Arena · LLM-as-Judge"),
        ("43", "PEFT & LoRA",              "低秩适配 · QLoRA · 10GB微调7B"),
        ("44", "Full FT vs PEFT",          "灾难性遗忘 · 多Adapter"),
        ("45", "偏好数据",                 "收集·多维度·偏差"),
        ("46", "对齐税",                   "能力退化 · 模式坍塌 · 平衡"),
        ("47", "RLHF 替代方案",            "Rejection · RRHF · SPIN"),
        ("48", "对齐的未来",               "Superalignment · 参与式对齐"),
    ]

    for num, title, desc in chapters:
        print(f"  {num:>2s} │ {title:<32s} │ {desc}")
    print()

    print(f"  运行示例:")
    print(f"    python3 33_sft/sft.py")
    print(f"    python3 38_dpo/dpo.py")
    print(f"    python3 43_peft_lora/peft_lora.py")
    print(f"    python3 48_alignment_future/alignment_future.py")

    print(f"\n  已完成: Part I (16) + Part II (16) + Part III (16) = 48/80")
    print(f"  仓库:   https://github.com/AgentZero2002/LLMAlignment")


if __name__ == "__main__":
    main()