The Llama 3 Herd of Models
本笔记基于摘要 + 公开资料,未读全文。
一句话讲什么(TL;DR)
Meta 把训练 Llama 3 大模型的全套"菜谱"公开了——用了什么料、多少张卡、跑多久、考多少分。
这是个什么场景 — 日常类比
想象你常去的米其林三星餐厅,平时只把成品端到桌上,菜谱、食材产地、火候温度一概不说。某天他突然把整本后厨工作手册甩出来:哪个农场的牛肉、几号灶台、几度烤几分钟、试菜请了多少评委、评委打了几分——一口气全摊给你看。Llama 3 这份报告就是这种级别的"全套菜谱"。市面上的对手是 GPT-4 / Claude 这类"只让你尝菜不让看后厨"的闭源餐厅;Meta 干脆把后厨大门打开,告诉你训一个前沿大模型到底要烧掉多少本钱。

之前的人怎么做的 — 3-5 bullet
- 闭源派(GPT-4 / Gemini / Claude):只放 API 和有限技术报告,数据规模、算力、训练细节都藏着
- 早期 Llama(Llama 2):开源权重 + 较粗的报告,多模态能力缺失
- 其他开源基座(Mistral / Qwen / DeepSeek 早期版本):规模更小,或者只放权重不公开训练曲线
- 多模态接法(LLaVA / BLIP-2):在小语言模型上接视觉,但底座本身不是前沿规模
- 结果:开源社区缺一个"接近 GPT-4 级别 + 训练栈完全透明 + 自带视觉支路"的参考实现
这篇论文的关键想法
三件事一起做:
- 把规模拉到 405B:开源模型第一次正面冲击闭源 SOTA 量级,证明开源社区可以触及前沿
- 训练全栈透明:数据 pipeline、tokenizer、并行策略、训练损失曲线、failure recovery、scaling law 拟合,都写进报告
- 视觉适配器后挂:保留语言主干不动,把图像编码器通过 cross-attention 适配器接进去,避免重新训练破坏语言能力
核心立场是"规模 + 数据质量 + 工程稳定性 = 大部分能力",没有引入新的架构奇技淫巧(仍然是稠密 Transformer,没上 MoE)。

它怎么做的(方法)— 3-4 段
预训练数据 — 像采购食材:先去几百个网站抓原料,再过一遍质检流水线把烂菜叶挑出去。约 15T tokens(Llama 2 是 1.8T,扩了近 10 倍),多语言、代码、推理类样本占比上调。整套 pipeline 包括去重、质量分类器、毒性过滤、个人信息脱敏。比例怎么配也不是拍脑袋——先用小模型当替身(proxy)试不同搭配,效果好的那套再喂给大模型。
架构与 scaling — 像盖楼前先算钢筋用量:稠密 decoder-only Transformer,配上 GQA(Grouped-Query Attention,多人共享一份 KV 缓存)+ RoPE + SwiGLU,上下文 128K(先 8K 训完再扩展)。盖之前论文先拟合了一条 scaling law(规模与效果的经验曲线),用来反推 405B 在 15T tokens 下该停在哪、loss 该到多少。预训练动用了 16K 张 H100 GPU 量级,跑了数月(具体数字需读原文)。
等等,先慢一拍——稠密 Transformer 是什么?
稠密(dense)= 每过一次模型,所有参数都要参与计算;与之相对的 MoE(专家混合)= 每次只激活其中一小部分专家,省算力。Llama 3 选了"老老实实全员上场"这条路。
后训练(post-training)— 像反复试菜调味:先 SFT(教它说人话),再用 DPO(Direct Preference Optimization,直接告诉它"这个回答比那个好")配上拒绝采样(生成 N 个候选挑最好),来回 6 轮左右。没用更复杂的 PPO(强化学习那套),因为 DPO 更稳更便宜。
多模态适配器 — 像在主菜上加配菜:语言主干这道主菜不动,旁边接一个图像 encoder(ViT 类)+ 一组 cross-attention 层(让语言模型能"看见"图像 token)。分阶段训练:先冻住主干只训配菜部分,再联合微调。视频和语音也走同样的挂载思路,一个语言主干长出多条感知支路。
实验在做什么
- 基础语言评测:MMLU / GSM8K / HumanEval / MATH 等,405B 对标 GPT-4,70B 对标 GPT-3.5 / Claude Haiku 量级(具体数字需读原文)
- 长上下文:128K 上的 needle-in-a-haystack 类大海捞针测试
- 多语言:8 种主要语言的评测对比
- 代码与推理:分代码生成、debug、数学推理多个子任务
- 多模态:图像问答(VQA)、文档理解、图表解读、视频问答
- 安全与红队:jailbreak 抵抗、有害内容生成率、refuse rate 平衡
- 人类偏好:Arena 类盲测,看实际对话偏好胜率
你应该懂的几个新词 — 4-6 个
- GQA(Grouped-Query Attention):注意力的中间方案,多个 query head 共享一组 key/value head,省 KV cache。日常类比:一群学生(query)共用一份课本(kv),不用人手一本
- DPO(Direct Preference Optimization):偏好对齐方法,给一对回答(好 vs 坏)直接优化模型,不用先训 reward model 再 RL。比 PPO 简单一截
- 拒绝采样(Rejection Sampling):让模型生成 N 个候选,用判别器/奖励模型挑最好那个加进训练集,相当于自己给自己出"优等生答案"
- Cross-attention 适配器:在已有 Transformer 层之间插入新的注意力层,让外部信息(如图像 token)能"被看见",而不动原始主干权重
- Scaling Law:参数量、数据量、算力之间的经验幂律关系,用来在小规模拟合曲线后,预测大规模该停在哪
- Data mixing:训练时不同来源(网页/代码/书/多语言)按什么比例喂入,比例选错性能差异巨大
它和其他论文什么关系
- 承接 Llama 2(2023):同家族升级,规模 ×10,加多模态分支
- 对标闭源前沿:GPT-4(OpenAI)、Gemini 1.5(Google)、Claude 3(Anthropic)——同一档位的稠密大模型
- 对比 MoE 路线:Mixtral / DeepSeek-V2 / Qwen-MoE 走稀疏激活,Llama 3 坚持稠密
- 后被引用:成为 2024-2025 开源基座事实标准,很多 RLHF / agent / VLM 工作直接 finetune Llama 3
- 多模态思路相关:Flamingo(cross-attention 视觉适配器祖师爷)、LLaVA(投影层接法)、BLIP-2(Q-Former),Llama 3 视觉支路接近 Flamingo 派
- 训练栈透明度对标:BLOOM 报告、OPT 报告、GPT-NeoX 报告——但 Llama 3 是第一份"前沿规模 + 全栈细节"的开源报告
我建议这样读 — 3-4 步
- 先读 §1 + §2 + §10(结论):搞清楚他们想证明什么、最后证明到了什么
- 再读 §3 数据 pipeline + §5 预训练:这是工程含金量最高、最值得抄作业的部分
- 跳到 §7 后训练(DPO + 拒绝采样的迭代循环):理解 SFT 之后到底是怎么把模型调"听话"的
- 多模态部分(§8)单独对照 Flamingo / LLaVA 看:把它当成"视觉适配器的工业实现案例",而不是新架构
如果只看 30 分钟:读 §1、§5.1(数据)、§7(后训练循环图)、§9(评测表)就够。
为什么值得读
- 行业基线手册:要做大模型训练,这是 2024 年最权威的"应该怎么做"参考,回避了一堆隐性陷阱
- 工程透明度天花板:从 tokenizer 到 failure recovery 都写出来了,对工程同学的价值远超论文本身
- 多模态接法的工业模板:报告里的"主干冻结 + 适配器后挂 + 分阶段联合训"是后续 VLM / 视频/ 语音模型反复用的范式
- 理解开源生态:Llama 3 是 2024-2025 年 fine-tune / agent / 具身智能上层应用的事实底座,下游论文几乎都建在它上面,读了它才知道下游论文的"地基"长什么样
- Scaling law 实战:工业上真把 scaling law 用到 405B 这种规模并把过程写出来,对学习"如何决定下一个模型多大"非常有价值
◼
引用本笔记 / Cite this note
@online{eai_llama_3_herd_2026,
title = {(readable note) The Llama 3 Herd of Models},
author = {Zhou, Jason},
year = {2026},
note = {Note on a 2024 paper},
howpublished = {\url{https://estelledc.github.io/embodied-ai-reading-station/papers/llama-3-herd/}},
organization = {Embodied AI Reading Station}
}
All 156 papers (full index)
- 1. LLaVA: Visual Instruction Tuning
- 2. 3DShape2VecSet: 3D Shape Representation for Diffusion Models
- 3. SayCan: Do As I Can, Not As I Say
- 4. OpenVLA: An Open-Source Vision-Language-Action Model
- 5. VLAS: VLA Model With Speech Instructions
- 6. MLA: Multisensory Language-Action Model
- 7. Cosmos Policy: Fine-Tuning Video Models for Visuomotor Control
- 8. CartoRadar: RF-Based 3D SLAM Rivaling Vision Approaches
- 9. mmCLIP: Boosting mmWave-based Zero-shot HAR via Signal-Text Alignment
- 10. mmNorm: Non-Line-of-Sight 3D Object Reconstruction via mmWave Surface Normal Estimation
- 11. Proactive Hearing Assistants that Isolate Egocentric Conversations
- 12. NeuralAids: Wireless Hearables With Programmable Speech AI Accelerators
- 13. Creating speech zones with self-distributing acoustic swarms
- 14. Conv-TasNet: Surpassing Ideal Time-Frequency Magnitude Masking for Speech Separation
- 15. SoundStream: An End-to-End Neural Audio Codec
- 16. AudioLM
- 17. Conformer
- 18. Dual-path RNN
- 19. EnCodec
- 20. Meta-StyleSpeech
- 21. MusicLM
- 22. Robust Speech Recognition via Large-Scale Weak Supervision
- 23. SeamlessM4T
- 24. Stable Audio
- 25. Universal Source Separation with Weakly Labelled Data
- 26. Meta-World: A Benchmark and Evaluation for Multi-Task and Meta Reinforcement Learning
- 27. RLBench: The Robot Learning Benchmark & Learning Environment
- 28. robosuite: A Modular Simulation Framework and Benchmark for Robot Learning
- 29. BridgeData V2
- 30. CALVIN
- 31. LIBERO
- 32. RH20T
- 33. What Matters in Learning from Offline Human Demonstrations for Robot Manipulation
- 34. DROID
- 35. Open X-Embodiment
- 36. RoboCasa
- 37. SimplerEnv
- 38. Diffusion Policy: Visuomotor Policy Learning via Action Diffusion
- 39. 3D Diffusion Policy: Generalizable Visuomotor Policy Learning via Simple 3D Representations
- 40. Consistency Policy: Accelerated Visuomotor Policies via Consistency Distillation
- 41. EquiBot: SIM(3)-Equivariant Diffusion Policy
- 42. DiT-Policy
- 43. Diffusion Policy Policy Optimization (DPPO)
- 44. Affordance-based Robot Manipulation with Flow Matching
- 45. FlowPolicy: 3D Flow-based Policy via Consistency Flow Matching
- 46. FAST: Efficient Action Tokenization for VLA
- 47. pi_0: Vision-Language-Action Flow Model
- 48. pi_0.5: VLA with Open-World Generalization
- 49. A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning
- 50. Generative Adversarial Imitation Learning
- 51. Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware (ACT/ALOHA)
- 52. AnyTeleop
- 53. Behavior Transformers: Cloning k Modes with One Stone
- 54. Implicit Behavioral Cloning
- 55. RoboCat
- 56. ALOHA 2
- 57. DexCap
- 58. HumanPlus
- 59. Generalizable Humanoid Manipulation with 3D Diffusion Policies (iDP3)
- 60. Mobile ALOHA
- 61. SmolVLA
- 62. Universal Manipulation Interface
- 63. Behavior Generation with Latent Actions (VQ-BeT)
- 64. ImageBind: One Embedding Space To Bind Them All
- 65. Connecting Touch and Vision via Cross-Modal Prediction
- 66. AnyMAL: An Efficient and Scalable Any-Modality Augmented Language Model
- 67. AudioPaLM
- 68. FROMAGe: Grounding LLMs to Images
- 69. OneLLM
- 70. X-VLM: Multi-Grained Vision Language Pre-Training
- 71. Tactile Beyond Pixels (Sparsh-X)
- 72. Sparsh: Self-supervised Touch Representations
- 73. Tactile-VLA
- 74. TLA: Tactile-Language-Action
- 75. Code as Policies: Language Model Programs for Embodied Control
- 76. Inner Monologue: Embodied Reasoning through Planning with Language Models
- 77. LLM+P: Empowering LLMs with Optimal Planning
- 78. PaLM-E: An Embodied Multimodal Language Model
- 79. ProgPrompt
- 80. ChatGPT for Robotics
- 81. GenSim
- 82. RoboFlamingo
- 83. Tree-Planner
- 84. VoxPoser
- 85. See Through Smoke: Robust Indoor Mapping with Low-cost mmWave Radar
- 86. Can WiFi Estimate Person Pose?
- 87. 3DRIMR: 3D Reconstruction and Imaging via mmWave Radar based on Deep Learning
- 88. milliEgo: Single-chip mmWave Radar Aided Egomotion Estimation via Deep Sensor Fusion
- 89. High Resolution Point Clouds from mmWave Radar
- 90. RadarSLAM: Radar based Large-Scale SLAM in All Weathers
- 91. Through-Wall Pose Imaging in Real-Time with a Many-to-Many Encoder/Decoder Paradigm
- 92. RFMask: A Simple Baseline for Human Silhouette Segmentation with Radio Signals
- 93. RFPose-OT: RF-Based 3D Human Pose Estimation via Optimal Transport Theory
- 94. Argus: Multi-View Egocentric Human Mesh Reconstruction Based on Stripped-Down Wearable mmWave Add-on
- 95. Diffusion Model is a Good Pose Estimator from 3D RF-Vision
- 96. Enabling Visual Recognition at Radio Frequency (PanoRadar)
- 97. Wave-Former: Through-Occlusion 3D Reconstruction via Wireless Shape Completion
- 98. Habitat: A Platform for Embodied AI Research
- 99. Isaac Gym: High Performance GPU-Based Physics Simulation For Robot Learning
- 100. DexMV
- 101. Habitat 2.0
- 102. ManiSkill
- 103. ProcTHOR
- 104. SAPIEN: A SimulAted Part-based Interactive ENvironment
- 105. BEHAVIOR-1K
- 106. Habitat 3.0
- 107. Isaac Lab
- 108. MuJoCo Playground
- 109. RT-1: Robotics Transformer for Real-World Control at Scale
- 110. 3D Diffusion Policy (DP3)
- 111. Octo: An Open-Source Generalist Robot Policy
- 112. RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control
- 113. RT-Trajectory: Robotic Task Generalization via Hindsight Trajectory Sketches
- 114. 3D-VLA
- 115. DexVLA
- 116. GR-2: Generative Video-Language-Action Model
- 117. OpenHelix
- 118. OpenVLA-OFT
- 119. RDT-1B: Diffusion Foundation Model for Bimanual Manipulation
- 120. RoboMamba
- 121. SpatialVLA
- 122. TinyVLA
- 123. TraceVLA: Visual Trace Prompting
- 124. Learning Transferable Visual Models From Natural Language Supervision
- 125. Flamingo: a Visual Language Model for Few-Shot Learning
- 126. BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models
- 127. BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation
- 128. DeepSeek-VL: Towards Real-World Vision-Language Understanding
- 129. EVA-CLIP: Improved Training Techniques for CLIP at Scale
- 130. FILIP: Fine-grained Interactive Language-Image Pre-Training
- 131. Florence-2: Advancing a Unified Representation for a Variety of Vision Tasks
- 132. InternVL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks
- 133. Improved Baselines with Visual Instruction Tuning
- 134. OBELICS
- 135. Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond
- 136. Sigmoid Loss for Language Image Pre-Training
- 137. What matters when building vision-language models?
- 138. Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling
- 139. The Llama 3 Herd of Models
- 140. LLaVA-NeXT-Interleave
- 141. LLaVA-OneVision: Easy Visual Task Transfer
- 142. Long-CLIP: Unlocking the Long-Text Capability of CLIP
- 143. Pixtral 12B
- 144. Dream to Control: Learning Behaviors by Latent Imagination
- 145. World Models
- 146. DayDreamer
- 147. Mastering Atari with Discrete World Models
- 148. Dreamer V3: Mastering Diverse Domains through World Models
- 149. Transformers are Sample-Efficient World Models
- 150. TWM: Transformer-based World Models
- 151. 1X World Model Challenge
- 152. Cosmos World Foundation Model Platform
- 153. GAIA-1
- 154. Genie: Generative Interactive Environments
- 155. Navigation World Models
- 156. UniSim