Token Lab検索

READ → PREDICT → EXPERIMENT → EXPLAIN

学習ロードマップ

A connected curriculum for AI engineers and PhD students. Start at your prerequisite gap; complete the evidence checkpoint before moving on.

Detailed technical articles are available in English and Chinese. This locale shows the English article with localized navigation.

AI Engineer · Prioritize stages 2–6, then evaluate at stage 7. Deliver one working pipeline with cost and failure analysis.

PhD / Research · Start with stages 1 and 7, then select a paper track. Deliver an ablation and a reproducible negative or positive result.

  1. 1

    01 · Foundations

    Understand gradients, cross-entropy, KL and numerical stability before tuning models.

    Gradient debugger / 梯度检查 →Cross-entropy & KL →Adam →

    Checkpoint: explain one mechanism, run or specify one controlled test, and record one limitation.

  2. 2

    02 · LLM architecture

    Tokenize a multilingual input, verify a causal mask, then account for KV and attention memory.

    Tokenizer / 分词器 →Transformer →FlashAttention →

    Checkpoint: explain one mechanism, run or specify one controlled test, and record one limitation.

  3. 3

    03 · Representations & RAG

    Connect contrastive objectives to retrieval quality; compare chunking, hybrid search and reranking.

    CPC →SimCLR →Chunking / 分块 →Hybrid retrieval / 混合检索 →

    Checkpoint: explain one mechanism, run or specify one controlled test, and record one limitation.

  4. 4

    04 · Adaptation & training

    Choose the trainable parameters, budget memory and distinguish training loss from generalization.

    LoRA →QLoRA →AdamW →Train a tiny network / 训练小网络 →

    Checkpoint: explain one mechanism, run or specify one controlled test, and record one limitation.

  5. 5

    05 · RL & alignment

    Study MDPs and policy gradients, then PPO, RLHF, DPO and verifiable-reward reasoning.

    8-chapter RL path / RL 八章路线 →DPO →DeepSeek-R1 →

    Checkpoint: explain one mechanism, run or specify one controlled test, and record one limitation.

  6. 6

    06 · Agents & inference

    Build a bounded tool loop and compare cache, routing and quantization trade-offs.

    ReAct →Agent workflow →Model routing →Quantization →

    Checkpoint: explain one mechanism, run or specify one controlled test, and record one limitation.

  7. 7

    07 · Evaluation & research

    Audit judges, calibrate confidence and compare paired runs with fair search budgets.

    LLM judge →Calibration →Search budgets →Reproduction protocol →

    Checkpoint: explain one mechanism, run or specify one controlled test, and record one limitation.

  8. 8

    08 · Explain & interview

    Answer before revealing the analysis. Defend assumptions, edge cases and experimental evidence.

    500 original prompts / 500 原创引导题 →Reported interviews / 公开面经 →PhD exercises / 研究练习 →

    Checkpoint: explain one mechanism, run or specify one controlled test, and record one limitation.

OFFICIAL / EXTENDED LEARNING

Continue with official AI education

External official resources, checked 25 October 2026. Independent site; no endorsement or affiliation. These links leave Token Lab and follow the destination’s privacy and sign-in policies.