研究学习者路线The Researcher's Route
约 115 分钟about 115 minutes
将原理词条与论文卡交替阅读。每篇都记录研究问题、证据设置和适用边界；未披露的部分保留为未知。Alternate concepts and paper cards. Record each research question, evidence setup and scope; leave undisclosed details unknown.
- TransformerTransformer
- 注意力机制Attention
- 论文卡-Attention Is All You NeedPaper Card - Attention Is All You Need
- 预训练Pre-training
- 训练三阶段SFT → Reward Model → RL
- 论文卡-GPT-3Paper Card - GPT-3
- 缩放定律Scaling Laws
- 论文卡-ChinchillaPaper Card - Chinchilla
- 微调Fine-tuning
- 论文卡-InstructGPTPaper Card - InstructGPT
- 论文卡-Self-InstructPaper Card - Self-Instruct
- 论文卡-LIMAPaper Card - LIMA
- 知识蒸馏Knowledge Distillation
- 量化Quantization
- 困惑度Perplexity (PPL)
- 怎么读一篇AI论文How to Read an AI Paper
- 基准测试Benchmark
- 论文卡-HELMPaper Card - HELM
- 模型评测的局限Evaluation Limits
- 数据污染Data Contamination
- 开放权重Open Weights
- 论文卡-Llama 2Paper Card - Llama 2
- 模型卡Model Card
- 把一份技术报告的训练章节拆成预训练/后训练两段，各说出关键变量
- 看到「刷新 SOTA」的说法，先查评测集发布时间和训练截止日期的关系
- 读模型卡时能指出许可证、评测口径和已知局限三处关键信息
- 对路线里的论文卡，各说出研究问题与核心证据位置，并说出至少一条「论文没有主张」的流行说法
- Split a tech report's training section into pretraining and post-training, naming the key variables of each
- On seeing a "new SOTA" claim, first check the benchmark's release date against the model's training cutoff
- When reading a model card, point out the license, the eval setup, and the stated limitations
- For each paper card on this route, state its research question and where the core evidence sits, and name at least one popular claim the paper never makes
先试着回答，再展开参考答案。答不上来的地方，可从上方步骤返回对应词条。参考回答是核对要点，不要求逐字复述。Try answering before opening the reference. Revisit the relevant step above if needed. These answers highlight key ideas; you do not need to repeat the wording.
第 1 题 · 参考回答Question 1 · Reference answer
区分基础训练、监督数据和偏好优化，记录数据来源、目标函数与评测设置；技术报告没有披露的配方应标为未知，不按常见流程补写。Separate base training, supervised data and preference optimization. Record data sources, objectives and evaluation setup; leave undisclosed recipe details unknown.
第 2 题 · 参考回答Question 2 · Reference answer
对照题集公开时间、训练截止、基准版本与去重说明。时间重叠只能提示污染风险，不能单独证明泄题；还需看污染检查和未见任务表现。Compare dataset release, training cutoff, benchmark version and decontamination. Temporal overlap suggests risk, not proof of contamination; inspect checks and held-out tasks.
第 3 题 · 参考回答Question 3 · Reference answer
分别记录有实验支持、设置受限与未披露的项目。例如，模型卡只报告英文结果时，不推断中文表现或生产可靠性。Separate experimentally supported, setup-limited and undisclosed claims. English-only model-card results do not establish Chinese performance or production reliability.
第 4 题 · 参考回答Question 4 · Reference answer
示例：LIMA 研究少量高质量指令数据的作用，查看其受控人评与消融；不能据此声称少量数据能解决所有安全与鲁棒性问题。其他论文同样对应问题、证据和适用边界。For LIMA, inspect controlled human evaluation and ablations on small high-quality instruction datasets. This does not establish that little data solves every safety or robustness problem. Apply the same question/evidence/limits structure to other papers.
用在一个具体任务里Apply this to a concrete task
选一篇论文完成问题、证据与局限记录Read a paper into a question/evidence/limits record →
半小时入门The Half-Hour Primer · 一个下午看懂原理One Afternoon Under the Hood · 一周路线The Full Week · 内容工作者路线The Writer's Route · 动手开发者路线The Builder's Route · 选型决策者路线The Decision-Maker's Route · 运维上线路线The Operator's Route · 普通消费者路线The Everyday User's Route · 图像 / 视频 / 音频创作者路线The Image / Video / Audio Creator's Route