跳到主要内容

LLMs 占位文档补全清单

已完成（2026-06-04）：138 篇占位正文已全部替换为深度笔记。下文保留批次索引供查阅。

B1 — 预训练数据（5）✅​

  • 03-pre-training/01-pretraining-data/01-data-sources.md
  • 03-pre-training/01-pretraining-data/02-cleaning-deduplication.md
  • 03-pre-training/01-pretraining-data/03-quality-filtering.md
  • 03-pre-training/01-pretraining-data/04-data-mixture.md
  • 03-pre-training/01-pretraining-data/05-data-licensing.md

B2 — 分词 + 预训练目标（11）​

  • 03-pre-training/02-tokenization/01-tokenization-levels.md
  • 03-pre-training/02-tokenization/02-bpe.md
  • 03-pre-training/02-tokenization/03-wordpiece.md
  • 03-pre-training/02-tokenization/04-sentencepiece-unigram.md
  • 03-pre-training/02-tokenization/05-byte-level-bpe-tiktoken.md
  • 03-pre-training/02-tokenization/06-multilingual-tokenization.md
  • 03-pre-training/03-pretraining-objectives/01-causal-lm.md
  • 03-pre-training/03-pretraining-objectives/02-masked-lm.md
  • 03-pre-training/03-pretraining-objectives/03-prefix-lm-span-corruption.md
  • 03-pre-training/03-pretraining-objectives/04-fim.md
  • 03-pre-training/03-pretraining-objectives/05-multitask-pretraining.md

B3 — Scaling + 分布式（12）​

  • 03-pre-training/04-scaling-laws/01-kaplan-scaling-laws.md
  • 03-pre-training/04-scaling-laws/02-chinchilla-scaling-laws.md
  • 03-pre-training/04-scaling-laws/03-compute-vs-inference-optimal.md
  • 03-pre-training/04-scaling-laws/04-data-parameter-tradeoff.md
  • 03-pre-training/04-scaling-laws/05-emergent-abilities.md
  • 03-pre-training/05-distributed-training/01-data-parallelism.md
  • 03-pre-training/05-distributed-training/02-tensor-parallelism.md
  • 03-pre-training/05-distributed-training/03-pipeline-parallelism.md
  • 03-pre-training/05-distributed-training/04-zero-deepspeed.md
  • 03-pre-training/05-distributed-training/05-three-d-sequence-parallelism.md
  • 03-pre-training/05-distributed-training/06-fsdp.md
  • 03-pre-training/05-distributed-training/07-communication-optimization.md

B4 — 训练稳定性（5）​

  • 03-pre-training/06-training-stability/01-mixed-precision.md
  • 03-pre-training/06-training-stability/02-gradient-accumulation-clipping.md
  • 03-pre-training/06-training-stability/03-checkpointing-recomputation.md
  • 03-pre-training/06-training-stability/04-divergence-diagnosis.md
  • 03-pre-training/06-training-stability/05-loss-spike.md

B5 — SFT / 指令 / RLHF（17）​

  • 04-post-training-alignment/01-sft/01-sft-overview.md
  • 04-post-training-alignment/01-sft/02-data-construction.md
  • 04-post-training-alignment/01-sft/03-quality-quantity-tradeoff.md
  • 04-post-training-alignment/01-sft/04-catastrophic-forgetting.md
  • 04-post-training-alignment/02-instruction-tuning/01-flan-t0-self-instruct.md
  • 04-post-training-alignment/02-instruction-tuning/02-alpaca-vicuna-wizardlm.md
  • 04-post-training-alignment/02-instruction-tuning/03-high-quality-instruction-data.md
  • 04-post-training-alignment/03-rlhf/01-rlhf-pipeline.md
  • 04-post-training-alignment/03-rlhf/02-reward-model.md
  • 04-post-training-alignment/03-rlhf/03-ppo.md
  • 04-post-training-alignment/03-rlhf/04-kl-penalty-stability.md
  • 04-post-training-alignment/03-rlhf/05-rlhf-challenges.md

B6 — DPO / CAI / PEFT（13）​

  • 04-post-training-alignment/04-preference-optimization/01-dpo.md
  • 04-post-training-alignment/04-preference-optimization/02-ipo-kto-orpo-simpo.md
  • 04-post-training-alignment/04-preference-optimization/03-offline-vs-online.md
  • 04-post-training-alignment/04-preference-optimization/04-methods-comparison.md
  • 04-post-training-alignment/05-constitutional-ai-rlaif/01-constitutional-ai.md
  • 04-post-training-alignment/05-constitutional-ai-rlaif/02-rlaif.md
  • 04-post-training-alignment/05-constitutional-ai-rlaif/03-self-improvement-critique.md
  • 04-post-training-alignment/06-peft/01-adapter.md
  • 04-post-training-alignment/06-peft/02-prefix-prompt-p-tuning.md
  • 04-post-training-alignment/06-peft/03-lora-qlora.md
  • 04-post-training-alignment/06-peft/04-dora-lora-plus.md
  • 04-post-training-alignment/06-peft/05-peft-selection-guide.md

B7 — 推理部署（23）​

见 rg -l "正文由大纲自动补全生成" llms/05-inference-deployment

B8 — 推理能力 + 评估（22）​

见 llms/06-reasoning-test-time-compute 与 llms/07-evaluation

B9 — 技术报告占位（11）​

见 llms/08-technical-reports（排除已 rich 的 K2、GLM-4.6、V3.2、gpt-oss）

B10 — 前沿（18）​

见 llms/09-frontier-future（排除 01-mamba-ssm.md）

B11 — 附录（7）​

见 llms/10-appendix