主题
参考资料索引
以下按章节收录原始链接。每项主张与来源的对应关系、固定版本、实验条件和未验证范围，以正文为准。
第 01 章 先把语言任务写成可检查的输入和输出
- https://github.com/datawhalechina/happy-llm/blob/1e0b3a1e6f8936b8ba4bbd0673e8e2227d196c96/docs/chapter1/第一章 NLP基础概念.md
- https://arxiv.org/abs/1301.3781
- https://arxiv.org/abs/1802.05365
第 02 章 注意力怎样把信息搬到需要它的位置
- https://arxiv.org/abs/1409.0473
- https://github.com/datawhalechina/happy-llm/tree/1e0b3a1e6f8936b8ba4bbd0673e8e2227d196c96/docs/chapter2
- https://arxiv.org/abs/1706.03762
第 03 章 预训练目标决定模型能看见什么
- https://github.com/datawhalechina/happy-llm/tree/1e0b3a1e6f8936b8ba4bbd0673e8e2227d196c96/docs/chapter3
- https://arxiv.org/abs/1810.04805
- https://arxiv.org/abs/1907.11692
- https://arxiv.org/abs/1909.11942
- https://arxiv.org/abs/1910.10683
- https://arxiv.org/abs/2005.14165
- https://arxiv.org/abs/2302.13971
- https://arxiv.org/abs/2103.10360
第 04 章 训练规模很大，结论仍要落到证据上
- https://github.com/datawhalechina/happy-llm/tree/1e0b3a1e6f8936b8ba4bbd0673e8e2227d196c96/docs/chapter4
- https://arxiv.org/abs/2005.14165
- https://arxiv.org/abs/2203.15556
- https://arxiv.org/abs/2304.15004
- https://arxiv.org/abs/2203.02155
- https://arxiv.org/abs/1910.02054
- https://arxiv.org/abs/1707.06347
第 05 章 把一个小型 LLaMA 风格解码器拆开
- https://github.com/datawhalechina/happy-llm/blob/1e0b3a1e6f8936b8ba4bbd0673e8e2227d196c96/docs/chapter5/code/k_model.py
- https://arxiv.org/abs/1910.07467
- https://arxiv.org/abs/2104.09864
- https://arxiv.org/abs/2305.13245
- https://arxiv.org/abs/2002.05202
第 06 章 Tokenizer 和标签决定模型实际学到什么
- https://github.com/datawhalechina/happy-llm/tree/1e0b3a1e6f8936b8ba4bbd0673e8e2227d196c96/docs/chapter5/code
- https://arxiv.org/abs/1508.07909
- https://arxiv.org/abs/1808.06226
第 07 章 训练闭环：从损失到可重载的模型
第 08 章 使用 Transformers，仍然要核对模型和数据的约定
第 09 章 少训练一些参数，究竟省了什么
- https://github.com/datawhalechina/happy-llm/tree/1e0b3a1e6f8936b8ba4bbd0673e8e2227d196c96/docs/chapter6
- https://arxiv.org/abs/2106.09685
- https://arxiv.org/abs/2305.14314
- https://arxiv.org/abs/2305.18290
- https://arxiv.org/abs/2402.01306
- https://arxiv.org/abs/2101.00190
第 10 章 评测要回答一个明确的问题
- https://github.com/datawhalechina/happy-llm/tree/1e0b3a1e6f8936b8ba4bbd0673e8e2227d196c96/docs/chapter7
- https://arxiv.org/abs/2009.03300
- https://arxiv.org/abs/2110.14168
- https://arxiv.org/abs/2103.03874
- https://arxiv.org/abs/2311.12022
- https://arxiv.org/abs/1905.07830
- https://arxiv.org/abs/2402.13718
第 11 章 RAG 先找证据，再检查回答有没有用对证据
- https://github.com/datawhalechina/happy-llm/tree/1e0b3a1e6f8936b8ba4bbd0673e8e2227d196c96/docs/chapter7/RAG
- https://arxiv.org/abs/2005.11401
第 12 章 Agent 是一个受约束的执行循环
- https://github.com/datawhalechina/happy-llm/tree/1e0b3a1e6f8936b8ba4bbd0673e8e2227d196c96/docs/chapter7/Agent
- https://arxiv.org/abs/2210.03629
第13章 GRPO：从一组答案中产生学习信号
- https://github.com/datawhalechina/happy-llm/blob/1e0b3a1e6f8936b8ba4bbd0673e8e2227d196c96/docs/chapter8/grpo/01-demo-sync.py
- https://arxiv.org/abs/2402.03300
- https://github.com/datawhalechina/happy-llm/blob/1e0b3a1e6f8936b8ba4bbd0673e8e2227d196c96/docs/chapter8/grpo/README.md
第14章 OPD：让教师评价学生走过的路径
- https://github.com/datawhalechina/happy-llm/blob/1e0b3a1e6f8936b8ba4bbd0673e8e2227d196c96/docs/chapter8/opd/01-demo-sync.py
- https://github.com/datawhalechina/happy-llm/blob/1e0b3a1e6f8936b8ba4bbd0673e8e2227d196c96/docs/chapter8/opd/README.md
第15章 Search-R1：把搜索结果放回可训练的轨迹
- https://arxiv.org/abs/2503.09516
- https://github.com/datawhalechina/happy-llm/tree/1e0b3a1e6f8936b8ba4bbd0673e8e2227d196c96/docs/chapter8/search-r1
- https://github.com/datawhalechina/happy-llm/blob/1e0b3a1e6f8936b8ba4bbd0673e8e2227d196c96/docs/chapter8/search-r1/README.md
第16章 ReTool：工具执行反馈、奖励与隔离边界
- https://arxiv.org/abs/2504.11536
- https://github.com/datawhalechina/happy-llm/tree/1e0b3a1e6f8936b8ba4bbd0673e8e2227d196c96/docs/chapter8/retool
- https://github.com/datawhalechina/happy-llm/blob/1e0b3a1e6f8936b8ba4bbd0673e8e2227d196c96/docs/chapter8/retool/README.md
第 17 章 小模型微调：先把一项任务做可靠
- https://github.com/datawhalechina/happy-llm/blob/1e0b3a1e6f8936b8ba4bbd0673e8e2227d196c96/Extra-Chapter/why-fine-tune-small-large-language-models/readme.md
- https://github.com/datawhalechina/happy-llm/blob/1e0b3a1e6f8936b8ba4bbd0673e8e2227d196c96/Extra-Chapter/why-fine-tune-small-large-language-models/why-fine-tune-small-large-language-models.ipynb
第 18 章 Transformer 设计复盘：从文本到注意力，哪里最容易接错
- https://github.com/datawhalechina/happy-llm/blob/1e0b3a1e6f8936b8ba4bbd0673e8e2227d196c96/Extra-Chapter/transformer-architecture/readme.md
- https://github.com/datawhalechina/happy-llm/blob/1e0b3a1e6f8936b8ba4bbd0673e8e2227d196c96/Extra-Chapter/text-data-processing/readme.md
- https://github.com/xinala-781
- https://github.com/pytorch/pytorch/blob/v2.5.1/torch/nn/modules/transformer.py
第 19 章 拼接式视觉语言模型：图像真的进入语言模型了吗
- https://github.com/datawhalechina/happy-llm/blob/1e0b3a1e6f8936b8ba4bbd0673e8e2227d196c96/Extra-Chapter/vlm-concatenation-finetune/README.md
- https://github.com/datawhalechina/happy-llm/tree/1e0b3a1e6f8936b8ba4bbd0673e8e2227d196c96/Extra-Chapter/vlm-concatenation-finetune
- https://github.com/ShaohonChen/Qwen3-SmVL
- https://arxiv.org/abs/2504.05299
第20章 思考预算与生成：更多 token 应该花在哪里
- https://arxiv.org/abs/2501.19393
- https://docs.vllm.ai/en/latest/api/vllm/sampling_params/
- https://github.com/datawhalechina/happy-llm/blob/1e0b3a1e6f8936b8ba4bbd0673e8e2227d196c96/Extra-Chapter/s1-vllm-thinking-budget/readme.md
第21章 CDDRS：把“找相似内容”变成可审计的知识检索
- https://github.com/datawhalechina/happy-llm/blob/1e0b3a1e6f8936b8ba4bbd0673e8e2227d196c96/Extra-Chapter/CDDRS/readme.md
- https://github.com/datawhalechina/happy-llm/blob/1e0b3a1e6f8936b8ba4bbd0673e8e2227d196c96/Extra-Chapter/CDDRS/CDDRS.ipynb