← Frontier Lab Interview Curriculum·Study guide 05 of 13
Senior-staff study guide · 05

LLM internals: architecture, pretraining, post-training

LLM internals covers how today's models are built: the Transformer and its modern variants, the scaling and data decisions behind pretraining, and the post-training stack (SFT, preference optimization, RL for reasoning) that turns a base model into a product. Frontier labs probe it because research-track and senior engineers are expected to reason about why their models behave the way they do and what to change. At senior staff the bar is not definitions but trade-offs with numbers (KV-cache bytes, FLOPs, token budgets), diagnosis of failures in real runs, and a defended critique of a specific lab's model with an experiment to fix it.

90resources11courses13with video44from 2025–2651%advanced48/48question groups covered2%of interview evidence (0% at staff+)
Intro 6Core 38Advanced 46

The senior-staff bar

What L7-level depth looks like in this topic, the follow-up an interviewer would push on, and the resources that prepare you for it.

From the interview reportsAt L7 research-track loops, expect 'what's wrong with our models and how would you fix it' (Meta GenAI E7 talk topic) and rapid-fire questions across the stack (positional encodings, 5D parallelism, PPO vs GRPO).

Start here

The best resources for a time-poor senior engineer, newest first.

2026Book

Reinforcement Learning from Human Feedback (RLHF Book)

Nathan Lambert · Sep 2026 · ~15 h

The best single text on post-training: instruction tuning, reward models, rejection sampling, PPO/GRPO, DPO-family methods, on-policy distillation, plus evaluation and over-optimization.

2026Course

CS336: Language Modeling from Scratch (Spring 2026)

Percy Liang, Tatsunori Hashimoto (Stanford) · Apr 2026 · ~70 h

The one course that walks the whole stack you'll be quizzed on: tokenizer, Transformer, GPU kernels, parallelism, scaling laws, data filtering, SFT and RL, each with a from-scratch assignment.

2025Practice

nanochat

Andrej Karpathy · Oct 2025 · ~12 h

Single small codebase that trains tokenizer, pretrains, mid-trains, SFTs and RLs a chat model for about $100 on one 8×H100 node (speedrun.sh).

2025Blog post

The Big LLM Architecture Comparison

Sebastian Raschka (Ahead of AI) · Jul 2025 · ~3 h

Side-by-side diagrams of DeepSeek-V3, OLMo 2, Gemma 3, Llama 4, Qwen3, Kimi K2, gpt-oss and others: MLA vs GQA, norm placement, QK-norm, MoE layout, sliding windows.

2024Paper

DeepSeek-V3 Technical Report

DeepSeek-AI · Dec 2024 · ~6 h

MLA, fine-grained MoE with shared experts, auxiliary-loss-free load balancing, multi-token prediction, FP8 training and DualPipe, with costs.

2024Paper

The Llama 3 Herd of Models

Llama Team, AI @ Meta · Jul 2024 · ~8 h

The most complete public account of a frontier dense-model pipeline: data mix and annealing, scaling-law-chosen size, 4D parallelism, reliability at 16K GPUs, and iterated SFT + rejection sampling + DPO.

Learning path

Work through the stages in order. The last stage is the senior-staff layer: production case studies, papers to critique and large-scale trade-offs. Each resource's L7 line says what it adds at your level.

  1. 1

    Foundations: get fluent fast

    Know the full pipeline and the vocabulary in about a week · Week 1 (8–10 h)

    • 2025
      Deep Dive into LLMs like ChatGPT
      Andrej Karpathy · Feb 2025 · Lecture · Intro · ~3.5 h · video
      Plain-language tour of the full pipeline: data, tokenization, pretraining, SFT, RLHF and RL for reasoning, with good intuition for hallucination and why models fail.
      L7Baseline; the fastest single overview if the post-training stack is new to you.
    • 2024
      Let's reproduce GPT-2 (124M)
      Andrej Karpathy · Jun 2024 · Lecture · Core · ~4 h · video
      Four hours from empty file to a trained GPT-2, including mixed precision, FlashAttention, learning-rate schedule, gradient accumulation and DDP.
      L7Lets you talk through the real training-loop details (warmup, weight decay groups, loss spikes) the way someone who has run one does.
      Let's build GPT: from scratch (2023)nanoGPT repo
    • 2024
      Let's build the GPT Tokenizer
      Andrej Karpathy · Feb 2024 · Lecture · Core · ~2 h · video
      Builds byte-level BPE from scratch and shows the failure modes (arithmetic, trailing spaces, glitch tokens, multilingual cost) that trace back to tokenization.
      L7Lets you explain model weaknesses that come from the tokenizer, a common thread in 'what's wrong with our model' questions.
    • 2017
      Attention Is All You Need
      Ashish Vaswani et al. (Google) · Jun 2017 · Paper · Core · ~2 h
      The original Transformer; read to know what has changed since (post-norm, sinusoidal positions, ReLU FFN, encoder-decoder).
      L7Baseline; skim if known, but be able to list every component modern models replaced and why.
    • 2019
      Language Models are Unsupervised Multitask Learners (GPT-2)
      Alec Radford et al. (OpenAI) · Feb 2019 · Paper · Core · ~1.5 h
      The decoder-only, pre-norm baseline every architecture tour starts from; also the WebText data idea.
      L7Baseline; the 'before' column of your architecture diff table.
      GPT-3: Language Models are Few-Shot Learners (2020)
    • 2022
      Training language models to follow instructions with human feedback (InstructGPT)
      Long Ouyang et al. (OpenAI) · Mar 2022 · Paper · Core · ~2 h
      The canonical SFT → reward model → PPO recipe with the KL penalty and pretraining mix, plus the labeler process.
      L7Baseline for every RLHF question; know the three stages and the alignment tax result cold.
    • 2025
      The Illustrated DeepSeek-R1
      Jay Alammar · Jan 2025 · Blog post · Intro · ~1 h
      Diagram-first explanation of the R1 pipeline: R1-Zero pure RL, cold-start SFT, reasoning RL, rejection sampling and distillation.
      L7Baseline; read before the R1 paper to keep the stages straight.
      The Illustrated GPT-OSS
  2. 2

    Architecture tour

    Build the GPT-2 → Llama 3 → DeepSeek-V3 → Qwen3 diff table and defend each choice · Weeks 2–3 (12–15 h)

    • 2025
      ★The Big LLM Architecture Comparison
      Sebastian Raschka (Ahead of AI) · Jul 2025 · Blog post · Core · ~3 h
      Side-by-side diagrams of DeepSeek-V3, OLMo 2, Gemma 3, Llama 4, Qwen3, Kimi K2, gpt-oss and others: MLA vs GQA, norm placement, QK-norm, MoE layout, sliding windows.
      L7The fastest way to build the curriculum's one-page architecture diff table and say why each lab chose what it did.
      LLM Architecture Gallery (kept up to date, incl. Qwen3-Next/3.5 hybrids)From GPT-2 to gpt-oss
    • 2023
      GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints
      Joshua Ainslie et al. (Google) · May 2023 · Paper · Core · ~1 h
      Short paper introducing grouped-query attention, the KV-cache saving now used in Llama, Qwen and most dense models.
      L7Lets you quantify KV-cache savings of MHA vs MQA vs GQA and the quality cost.
    • 2024
      DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model
      DeepSeek-AI · May 2024 · Paper · Advanced · ~3 h
      Introduces Multi-head Latent Attention (low-rank KV compression with decoupled RoPE); the clearest derivation of MLA.
      L7Lets you compare MLA vs GQA on KV-cache bytes per token and explain why decoupled RoPE is needed.
    • 2019
      Root Mean Square Layer Normalization
      Biao Zhang, Rico Sennrich · Oct 2019 · Paper · Core · ~1 h
      Short paper behind RMSNorm: drops mean-centring from LayerNorm for speed with little quality loss.
      L7Answers the 'RMSNorm vs LayerNorm' question directly; pair with pre-norm vs post-norm and QK-norm from r18.
      Shazeer: GLU Variants Improve Transformer (SwiGLU, 2020)
    • 2021
      RoFormer: Enhanced Transformer with Rotary Position Embedding
      Jianlin Su et al. · Apr 2021 · Paper · Core · ~1.5 h
      Introduces RoPE, the default positional scheme in nearly every current LLM; work through the rotation-matrix derivation.
      L7Lets you derive why RoPE gives relative position through a dot product and why it decays with distance.
      ALiBi: Train Short, Test Long (2021)
    • 2023
      YaRN: Efficient Context Window Extension of Large Language Models
      Bowen Peng, Jeffrey Quesnelle, Honglu Fan, Enrico Shippole · Sep 2023 · Paper · Advanced · ~1.5 h
      The RoPE-scaling method (frequency interpolation plus attention temperature) used by Qwen, DeepSeek and others to extend context after pretraining.
      L7Lets you explain how labs go from 4K to 128K+ context cheaply and what breaks at long range.
      Round and Round We Go! What makes RoPE useful? (2024)
    • 2021
      Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity
      William Fedus, Barret Zoph, Noam Shazeer (Google) · Jan 2021 · Paper · Core · ~2 h
      Top-1 routing, capacity factor, auxiliary load-balancing loss and the training-stability tricks that made sparse MoE work.
      L7Lets you explain where MoE instability and load imbalance come from before discussing DeepSeek's fixes.
      Shazeer et al.: Sparsely-Gated Mixture-of-Experts (2017)
    • 2024
      A Survey on Mixture of Experts in Large Language Models
      Weilin Cai, Juyong Jiang, Fan Wang, Jing Tang, Sunghun Kim, Jiayi Huang · Jul 2024 · Survey paper · Core · ~4 h
      Covers routing, load balancing, expert granularity, shared experts and system design for MoE LLMs.
      L7Gives the vocabulary for the Mistral-style 'MoE and expert parallelism' question; pair with DeepSeek-V3 for the current recipe.
    • 2025
      Qwen3 Technical Report
      Qwen Team (Alibaba) · May 2025 · Paper · Advanced · ~4 h
      Dense and MoE family with QK-norm, no shared experts, 36T-token pretraining in three stages, and a single model with thinking and non-thinking modes plus strong-to-weak distillation.
      L7Lets you contrast Qwen's and DeepSeek's MoE choices and discuss thinking-budget control as a product decision.
    • 2025
      gpt-oss-120b & gpt-oss-20b Model Card
      OpenAI · Aug 2025 · Paper · Advanced · ~2 h
      OpenAI's open-weight MoE models: alternating dense and sliding-window attention, attention sinks, MXFP4 weights, harmony format and adjustable reasoning effort.
      L7The only recent first-party OpenAI architecture disclosure; useful for an OpenAI-specific critique.
    • 2025
      A Comprehensive Survey on Long Context Language Modeling
      Jiaheng Liu et al. · Mar 2025 · Survey paper · Core · ~4 h
      Positional encodings and extrapolation, context-extension training, efficient attention, KV-cache management and long-context evaluation in one place.
      L7Lets you answer 'different ways of encoding position' and 'how would you get to 1M context' with a full menu and trade-offs.
    • 2023
      Mistral 7B
      Mistral AI · Oct 2023 · Paper · Core · ~1 h
      Short report combining grouped-query attention with sliding-window attention and a rolling KV cache.
      L7Lets you say what GQA and SWA each buy (KV memory, long-sequence cost) and what they give up (heads, direct long-range attention).
    • 2024
      Mixtral of Experts
      Mistral AI · Jan 2024 · Paper · Core · ~1 h
      Sparse MoE with 8 experts and top-2 routing: ~47B total parameters but ~13B active per token, with routing analysis.
      L7Directly answers the Mistral 'why does 47B run like 13B' question; know where the memory cost still lands.
  3. 3

    Scaling arithmetic and pretraining

    Estimate a 70B run on paper and explain why labs over-train · Week 4 (10 h)

  4. 4

    Post-training and reasoning RL, hands-on

    Explain PPO vs GRPO in two minutes, then run a small GRPO job on math · Weeks 5–7 (25–30 h)

  5. 5

    Senior-staff depth

    Critique each target lab's models and propose a fix, citing frontier-scale evidence · Weeks 8–9 (15–20 h)

Question coverage

Every question group collected for this topic from candidate reports, mapped to the resources that prepare you for it. 48 of 48 groups are covered, which is 100% of all report mentions. Hover an example to see the full question.

ReportsQuestion group and examplePrepare with
2RLHF vs DPOMistral 2Preference-based training (reward-model RLHF vs DPO)Direct Preference Optimization: Your Language Model is Secretly a Reward ModelReinforcement Learning from Human Feedback (RLHF Book)Llama 2: Open Foundation and Fine-Tuned Chat ModelsDPO derivation vs RM+PPO pipeline; RLHF book direct-alignment chapter
2transformer componentsAnthropic 2What are the key components of a Transformer model and why does each matter?Attention Is All You NeedLet's reproduce GPT-2 (124M)The Big LLM Architecture ComparisonOriginal components, then what modern models replaced and why
2context window size effectsOpenAI 2What is the effect of adjusting an LLM's context window size?Lost in the Middle: How Language Models Use Long ContextsA Comprehensive Survey on Long Context Language ModelingYaRN: Efficient Context Window Extension of Large Language ModelsUsable vs nominal context, KV/compute cost, extension methods
1positional encoding variantsUnspecified frontier 1What are some different ways of encoding positional information?RoFormer: Enhanced Transformer with Rotary Position EmbeddingA Comprehensive Survey on Long Context Language ModelingYaRN: Efficient Context Window Extension of Large Language ModelsAbsolute, relative, RoPE, ALiBi and scaling variants compared
1PPO vs GRPOUnspecified frontier 1What is the difference between PPO and GRPO?DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language ModelsProximal Policy Optimization AlgorithmsThe State of Reinforcement Learning for LLM ReasoningGRPO drops the critic for group-relative advantages; PPO clipping/GAE
1async RL post-trainingMeta MSL 1Round on post-training async RL (architect an asynchronous RL post-training system)Asynchronous RLHF: Faster and More Efficient Off-Policy RL for Language Modelsopen-instructOlmo 3Async generation/learning split, staleness tolerance, real RLVR infra
1mixture of expertsMistral 1Mixture of experts (and expert parallelism)A Survey on Mixture of Experts in Large Language ModelsDeepSeek-V3 Technical ReportSwitch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient SparsityRouting, load balancing, expert parallelism in DeepSeek-V3 and Switch
1RMSNorm vs LayerNormMistral 1Normalization (RMSNorm vs LayerNorm in transformer blocks)Root Mean Square Layer NormalizationThe Big LLM Architecture ComparisonRMSNorm math, pre-norm placement, QK-norm in current models
1novel data miningOpenAI 1Design an ML system to mine novel/interesting images from a huge unlabeled corpus, using human labelers without over-relying on themAutomatic Data Curation for Self-Supervised Learning: A Clustering-Based ApproachA Survey on Data Selection for Language ModelsEmbedding clustering for diverse mining; labeler budget via active selection
1transformer basicsOpenAI 1ML coding testing familiarity with the Transformer architecture and its basic conceptsLet's reproduce GPT-2 (124M)CS336: Language Modeling from Scratch (Spring 2026)Code a GPT from scratch; CS336 assignment 1 Transformer
1transformer design rationaleOpenAI 1Transformer internals and scaling laws, with emphasis on the "why" behind design choicesThe Big LLM Architecture ComparisonTraining Compute-Optimal Large Language Models (Chinchilla)How To Scale Your ModelWhy each design choice; scaling laws and Transformer FLOP math
1transformer internals depthGoogle DeepMind 1Deep discussion of NLP and transformer internalsCS336: Language Modeling from Scratch (Spring 2026)CMU 11-711 Advanced NLP (Spring 2026)Attention Is All You NeedLecture-level depth on attention, decoding and Transformer internals
1code RL algorithmsAnthropic 1Research Engineer tech screen focused on coding RL algorithms (per the recruiter's prep link)Spinning Up in Deep RLTRL GRPO TrainerCS336: Language Modeling from Scratch (Spring 2026)Implement REINFORCE/PPO, GRPO trainer, CS336 assignment 5 GRPO
1RL training failure modesUnspecified frontier 1RL training debugging scenarios: what do you do when you hit entropy collapse, the model reasoning in another language, or terrible MFU?The Entropy Mechanism of Reinforcement Learning for Reasoning Language ModelsDAPO: An Open-Source LLM Reinforcement Learning System at ScaleDeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement LearningHow To Scale Your ModelEntropy collapse fixes, R1 language-mixing reward, MFU roofline math
1attention scalingUnspecified frontier 1Why does attention divide by sqrt(d_k)?Attention Is All You NeedGoogle DeepMind Interview Notes: LLM Fundamentals and RL-for-LLMs quizzesSection 3.2.1 scaling rationale; quiz Q1 variance derivation
1transformer limitationsUnspecified frontier 1Name one drawback of transformer models.Efficient Transformers: A SurveyA Comprehensive Survey on Long Context Language ModelingQuadratic cost, KV memory, length generalization limits
1MQA vs GQA tradeoffsMeta MSL 1, Mistral 1What are Multi-Query Attention (MQA) and Grouped-Query Attention (GQA), and what do they trade away?GQA: Training Generalized Multi-Query Transformer Models from Multi-Head CheckpointsDeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language ModelMQA/GQA KV savings vs quality; MLA as alternative
1RoPE and context extensionMeta MSL 1Explain RoPE and how position interpolation / YaRN extend context beyond the trained length.RoFormer: Enhanced Transformer with Rotary Position EmbeddingYaRN: Efficient Context Window Extension of Large Language ModelsRoPE derivation; position interpolation and YaRN frequency scaling
1Chinchilla scaling lawsAnthropic 1What do the Chinchilla scaling laws say, and how do they differ from earlier scaling intuitions?Training Compute-Optimal Large Language Models (Chinchilla)Beyond Chinchilla-Optimal: Accounting for Inference in Language Model Scaling LawsChinchilla 20 tokens/param vs Kaplan; inference-aware over-training
1MoE capacity vs FLOPsMistral 1, Cohere 1What is a mixture-of-experts architecture and how does it scale capacity without scaling FLOPs?Mixtral of ExpertsSwitch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient SparsityA Survey on Mixture of Experts in Large Language ModelsTotal vs active parameters, sparse routing keeps FLOPs per token fixed
1pretraining vs SFT vs preferenceMeta MSL 1Explain the difference between pre-training, supervised fine-tuning and preference optimisation.Deep Dive into LLMs like ChatGPTTraining language models to follow instructions with human feedback (InstructGPT)Reinforcement Learning from Human Feedback (RLHF Book)Pipeline stages and what each objective optimizes
1decoding sampling strategies comparedGoogle DeepMind 1Compare greedy, beam search, top-k, top-p and temperature sampling. When does each fail?The Curious Case of Neural Text DegenerationGreedy/beam degeneration, top-k vs nucleus, temperature effects
1SwiGLU gated activationsMeta MSL 1Explain SwiGLU and why gated activations replaced ReLU/GELU in modern LLM MLP blocks.Root Mean Square Layer NormalizationThe Big LLM Architecture ComparisonGLU Variants paper (extra link) and SwiGLU in modern models
1decoder forward pass tensor shapesAnthropic 1Walk me through what happens, tensor by tensor, in one forward pass of a decoder-only transformer.Let's reproduce GPT-2 (124M)How To Scale Your ModelGoogle DeepMind Interview Notes: LLM Fundamentals and RL-for-LLMs quizzesTensor shapes through embedding, attention, MLP, logits; FLOP accounting
1catastrophic forgetting in fine-tuningMistral 2What is catastrophic forgetting and how do you mitigate it during fine-tuning?LoRA Learns Less and Forgets LessLoRA Without RegretLearning vs forgetting for LoRA and full FT; mitigation via rank, replay
Show the other 23 question groups
1Constitutional AI and RLAIFAnthropic 2What is Constitutional AI and how does it differ from RLHF? What is RLAIF?Constitutional AI: Harmlessness from AI FeedbackReinforcement Learning from Human Feedback (RLHF Book)Constitution-guided self-critique and RLAIF vs human-label RLHF
1vision-language model image inputMeta MSL 1How do vision-language models get images into an LLM: projector, cross-attention, or native tokens?Vision Language Models (Better, faster, stronger)Visual Instruction Tuning (LLaVA)Projector (LLaVA), cross-attention (Flamingo), early fusion (Chameleon)
1image to video multimodalMeta MSL 2What changes when you move from images to video?Qwen2.5-VL Technical ReportVision Language Models (Better, faster, stronger)Frame sampling, temporal MRoPE, video token budgets
1attention-free transformer architecturesAnthropic 1Explain attention-free transformer architectures and their trade-offs.Mamba: Linear-Time Sequence Modeling with Selective State SpacesEfficient Transformers: A SurveyGated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-FreeSSMs and linear attention vs softmax attention; hybrid trade-offs
1matrix ops in LLMsAnthropic 1Walk me through matrix manipulations relevant to LLM architectures.How To Scale Your ModelLet's reproduce GPT-2 (124M)Transformer matmul shapes, FLOPs and memory per operation
1self-attention complexity long contextOpenAI 1Explain self-attention. What is its computational complexity, and what are your options when contexts get long?Efficient Transformers: A SurveyA Comprehensive Survey on Long Context Language ModelingHow To Scale Your ModelO(T^2) cost, FlashAttention, sparse/linear attention, KV compression
1loss divergence long runGoogle DeepMind 1Your pretraining loss suddenly diverges at step 300k of a long run. Diagnose and fix it.Small-scale proxies for large-scale Transformer training instabilitiesOPT-175B training chronicles (logbook)Kimi K2: Open Agentic IntelligenceLogit growth, QK-norm, z-loss, rollback-and-skip practice
1Llama architecture choicesMeta MSL 1Explain the architectural choices in a Llama-class model: why grouped-query attention, RoPE and SwiGLU instead of the vanilla 2017 Transform…The Llama 3 Herd of ModelsThe Big LLM Architecture ComparisonGQA: Training Generalized Multi-Query Transformer Models from Multi-Head CheckpointsRoFormer: Enhanced Transformer with Rotary Position EmbeddingGQA, RoPE, SwiGLU, RMSNorm choices justified in Llama 3
1post-training recipe for assistantMeta MSL 1Walk me through a post-training recipe to turn a pretrained base model into a personalised assistant.Tulu 3: Pushing Frontiers in Open Language Model Post-TrainingThe Llama 3 Herd of ModelsReinforcement Learning from Human Feedback (RLHF Book)Tülu 3 and Llama 3 multi-stage SFT/DPO/RL recipes
1loss spike debuggingxAI 1Loss spikes mid-run on a large pretraining job. Walk me through your debugging process.Small-scale proxies for large-scale Transformer training instabilitiesOPT-175B training chronicles (logbook)The Smol Training Playbook: The Secrets to Building World-Class LLMsSmall-scale proxies, real logbook of spike handling, playbook failures
1streaming web-scale dedupxAI 1Design a deduplication pipeline for a web-scale pretraining corpus. It has to run as a streaming process.Deduplicating Training Data Makes Language Models BetterThe FineWeb Datasets: Decanting the Web for the Finest Text Data at ScaleMinHash and suffix-array dedup; FineWeb dedup ablations; datatrove
1GQA and sliding window attentionMistral 1Mistral 7B shipped with grouped-query attention and sliding-window attention. What does each buy you, and what does each cost?Mistral 7BGQA: Training Generalized Multi-Query Transformer Models from Multi-Head CheckpointsGQA KV savings and sliding-window rolling cache trade-offs
1Mixtral sparse MoE costMistral 1Explain how a Mixtral-style sparse mixture-of-experts model works. Why does a ~47B-parameter model run at roughly the cost of a ~13B one?Mixtral of ExpertsA Survey on Mixture of Experts in Large Language ModelsTop-2 of 8 experts: active vs total parameter accounting
1sparse MoE enterprise deploymentCohere 1Our flagship is a sparse MoE with ~10x more total than active parameters. Why is that architecture a good fit for private enterprise deploym…Mixtral of ExpertsA Survey on Mixture of Experts in Large Language ModelsDeepSeek-V3 Technical ReportHigh memory, low compute per token; serving and EP trade-offs
1domain adaptation of gen modelsAnthropic 1How would you adapt a generative AI model to perform well in a specific domain?LoRA Learns Less and Forgets LessAI Engineering (Ch. 7 Finetuning, Ch. 8 Dataset Engineering)Olmo 3Continued pretraining vs FT vs PEFT; mid-training data mixing
1sampling temperature creativityAnthropic 1How do you adjust the creativity of generative AI models?The Curious Case of Neural Text DegenerationTemperature and top-p control diversity vs coherence
1RLHF feedback pipelineOpenAI 1Design a feedback collection pipeline for RLHF.Llama 2: Open Foundation and Fine-Tuned Chat ModelsTraining language models to follow instructions with human feedback (InstructGPT)CS329H: Machine Learning from Human Preferences (Autumn 2026)Annotation protocol, rater agreement, reward-model data at scale
1loss divergence debuggingOpenAI 1Your training run's loss is diverging after step 10K. Walk me through your debugging process.Small-scale proxies for large-scale Transformer training instabilitiesOPT-175B training chronicles (logbook)Let's reproduce GPT-2 (124M)Divergence causes and checks: LR, logits, data, numerics
1MoE routing trade-offsMistral 1Discuss trade-offs in MoE routing: top-k vs expert-choice routing, load-balancing vs quality.Mixture-of-Experts with Expert Choice RoutingDeepSeek-V3 Technical ReportSwitch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient SparsityToken- vs expert-choice routing; aux-loss vs aux-free balancing
1X data ingestion pipelinexAI 1Design the data pipeline ingesting X content at petabyte scale for model training with safety filtering.The FineWeb Datasets: Decanting the Web for the Finest Text Data at ScaleA Survey on Data Selection for Language ModelsDeduplicating Training Data Makes Language Models BetterExtraction, quality/safety filters, dedup at petabyte scale
1LLM fundamentals quizGoogle DeepMind 1LLM Fundamentals: Attention, Transformers and the Training Pipeline (12 questions / 45–60 min)Google DeepMind Interview Notes: LLM Fundamentals and RL-for-LLMs quizzesCS336: Language Modeling from Scratch (Spring 2026)The quiz itself, backed by CS336 lectures
1RL for LLMs quizGoogle DeepMind 1RL for Language Models: Policy Gradients, PPO, GRPO and DPO (9 questions / 45–60 min)Google DeepMind Interview Notes: LLM Fundamentals and RL-for-LLMs quizzesReinforcement Learning from Human Feedback (RLHF Book)DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language ModelsRL quiz; RLHF book and GRPO derivations
1MoE expert parallelismMistral 1Explain a mixture-of-experts transformer and how expert parallelism distributes its work. Trace what happens to a batch of tokens from routi…DeepSeek-V3 Technical ReportA Survey on Mixture of Experts in Large Language ModelsSwitch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient SparsityAll-to-all dispatch, expert sharding, DualPipe overlap in DeepSeek-V3

All resources

Everything collected for this topic, newest first. 2026 items are marked green. ★ marks a top pick.

Type
Level
Subtopic

  • 2026
    CS329H: Machine Learning from Human Preferences (Autumn 2026)
    Sanmi Koyejo (Stanford) · Sep 2026 · Course · Advanced · ~15 h
    Covers the preference-learning theory behind reward models and DPO: discrete choice, Bradley–Terry, aggregation of disagreeing raters, and RL from preferences.
    L7Lets you discuss why reward models fail (rater noise, aggregation, Goodhart) from first principles rather than anecdotes.
    Running now; use the schedule and textbook links.
  • 2026
    ★Reinforcement Learning from Human Feedback (RLHF Book)
    Nathan Lambert · Sep 2026 · Book · Advanced · ~15 h
    The best single text on post-training: instruction tuning, reward models, rejection sampling, PPO/GRPO, DPO-family methods, on-policy distillation, plus evaluation and over-optimization.
    L7Lets you lay out a complete post-training recipe and say where each stage breaks, which is the core of 'how would you fix our model' answers.
    arXiv version (2504.12501, v12 Sep 2026)
  • 2026
    ★CS336: Language Modeling from Scratch (Spring 2026)
    Percy Liang, Tatsunori Hashimoto (Stanford) · Apr 2026 · Course · Advanced · ~70 h · video
    The one course that walks the whole stack you'll be quizzed on: tokenizer, Transformer, GPU kernels, parallelism, scaling laws, data filtering, SFT and RL, each with a from-scratch assignment.
    L7Lets you answer any 'why this design choice' question with something you have actually built and measured, not something you read.
    Assignments are heavy (each 15–25 h). Time-poor route: watch lectures on architecture, scaling, data and alignment; do assignments 1 and 5.
    YouTube playlist (Spring 2026)Assignment 1: basics (BPE, Transformer, AdamW)Assignment 3: scaling lawsAssignment 4: dataAssignment 5: alignment (SFT, GRPO on math)Spring 2025 offering
  • 2026
    CS25: Transformers United V6 (Spring 2026)
    Stanford CS25 organizers; guest speakers from frontier labs · Apr 2026 · Course · Advanced · ~12 h · video
    Weekly talks by the researchers building current models (architectures, reasoning, scaling, alternatives to attention); good for hearing how lab people frame open problems.
    L7Gives you current lab framing and names to cite when you lead an open-ended 'where is the architecture going' discussion.
    Seminar; pick talks by topic rather than watching all.
    All recordings (YouTube)
  • 2026
    DeepSeek-V4 (model card and technical report)
    DeepSeek-AI · Apr 2026 · Paper · Advanced · ~4 h
    Preview of V4-Pro (1.6T total / 49B active) and V4-Flash: hybrid compressed sparse attention for 1M context, manifold-constrained hyper-connections and Muon.
    L7Newest public frontier-scale design; lets you discuss where long-context architecture is heading with current numbers (KV cache and FLOPs vs V3.2).
    Technical report is linked from the model card; I verified the card, not a separate PDF URL.
  • 2026
    A Survey of Large Language Models
    Wayne Xin Zhao et al. (Renmin University) · Mar 2026 · Survey paper · Core · ~10 h
    Big, regularly revised map of the field (v19, Mar 2026): pretraining data, architectures, scaling, adaptation, utilization and evaluation, with tables of models.
    L7Reference to fill gaps; use its tables to check you have not missed a major technique.
    Very long; use as a lookup, not a read-through.
  • 2026
    CS224N: Natural Language Processing with Deep Learning (Winter 2026)
    Diyi Yang, Yejin Choi (Stanford); 2024 videos by Christopher Manning · Jan 2026 · Course · Intro · ~25 h · video
    Fastest structured route to NLP fluency for someone strong in ML but new to language: attention, Transformers, pretraining, post-training and evaluation lectures.
    L7Baseline; skim the Transformer, pretraining and post-training lectures only if the vocabulary is new.
    Winter 2026 videos are not public; the 2024 recording is the latest public set.
    YouTube playlist (Spring 2024, Manning)
  • 2026
    CS234: Reinforcement Learning (Winter 2026)
    Emma Brunskill (Stanford) · Jan 2026 · Course · Core · ~20 h · video
    Clean RL foundations (MDPs, policy gradients, actor-critic, offline RL, RLHF/DPO lectures) so PPO/GRPO stop being black boxes.
    L7Lets you derive the policy-gradient and advantage terms behind PPO and GRPO on a whiteboard instead of reciting them.
    2026 lectures are not public; use the 2024 recordings. Silver's 2015 course is older but still the canonical intro.
    YouTube playlist (Spring 2024)David Silver: UCL/DeepMind RL course (2015) videosDavid Silver teaching page (slides)
  • 2026
    CS285: Deep Reinforcement Learning (Spring 2026)
    Sergey Levine (UC Berkeley) · Jan 2026 · Course · Advanced · ~25 h · video
    Deeper RL than CS234: policy gradient variance, trust regions, actor-critic, offline RL and model-based RL, which is the theory under PPO-style LLM training.
    L7Lets you argue variance, bias and stability trade-offs (value baselines vs group baselines, KL vs clipping) at research depth.
    Spring 2026 recordings marked TBD on the site; Fall 2023 videos are the latest public set.
    YouTube playlist (Fall 2023 recordings)
  • 2026
    CMU 11-711 Advanced NLP (Spring 2026)
    Sean Welleck (CMU); adapted from Graham Neubig's Fall 2024 course · Jan 2026 · Course · Core · ~25 h · video
    Graduate NLP course centred on LLM foundations: modelling, decoding, pretraining, post-training, reasoning and RL, with code notebooks per lecture.
    L7Good second pass after CS336 for decoding and reasoning lectures; the code repo is quick to run against.
    I could not confirm public Spring 2026 videos; Neubig's Fall 2024 recordings are public.
    Lecture code (GitHub)Fall 2024 YouTube playlist (Neubig)Fall 2025 offeringNeubig Fall 2024 site
  • 2026
    Build a Reasoning Model (From Scratch)
    Sebastian Raschka (Manning) · 2026 · Book · Core · ~20 h · paid
    Code-first book that takes a small pretrained model and adds inference-time scaling, RL with verifiable rewards and distillation step by step.
    L7Gives you hands-on numbers for test-time compute and RLVR so you can defend claims about what reasoning training buys.
    Paid (Manning MEAP/print).
  • 2026
    Interconnects
    Nathan Lambert · 2026 · Blog post · Advanced · ~6 h
    Weekly analysis of open and closed model releases, post-training recipes and lab strategy from someone who trains open models at Ai2.
    L7Best source for forming the per-lab model critiques the curriculum asks for; read the posts on each target lab's latest release.
    Recent reasoning research: GRPO tweaks, base model RL, data curation
  • 2026
    TRL GRPO Trainer
    Hugging Face TRL · 2026 · Practice · Core · ~6 h
    Official docs and examples for running GRPO (and DAPO/Dr. GRPO loss variants) with custom reward functions; the quickest path to a small GRPO run on math.
    L7Lets you run the curriculum's GRPO experiment and report what the loss variants changed.
    open-r1 (full R1 reproduction)
  • 2026
    open-instruct
    Ai2 · 2026 · Practice · Advanced · ~6 h
    Ai2's production post-training code used for Tülu 3 and Olmo 3: SFT, DPO and RLVR with vLLM-backed generation.
    L7Reading real RLVR infrastructure lets you discuss generation/training overlap and off-policy staleness concretely.
  • 2026
    verifiers
    Prime Intellect · 2026 · Practice · Advanced · ~4 h
    Library for building RL environments and verifiable reward functions (single- and multi-turn, tool use) that plug into RL trainers.
    L7Lets you talk about reward and environment design for agentic RL, where much current lab work sits.
  • 2026
    Google DeepMind Interview Notes: LLM Fundamentals and RL-for-LLMs quizzes
    ShellyYin96 (GitHub, community) · 2026 · Practice · Core · ~4 h
    Two derivation-style oral quizzes (attention variance, KV-cache bytes, RoPE, scaling laws; policy gradients, PPO, GRPO, DPO) matching reported DeepMind rounds.
    L7Use as a timed self-test after the architecture and post-training stages; aim to answer with numbers in under 4 minutes each.
    Community repo, not an official DeepMind source.
    RL for Language Models quiz
  • 2025
    Olmo 3
    Ai2 (Allen Institute for AI) · Dec 2025 · Paper · Advanced · ~6 h
    Fully open model flow: every dataset, checkpoint, mid-training stage, SFT/DPO/RLVR recipe and the RL infrastructure, so you can see all the decisions a closed report omits.
    L7Lets you cite concrete ablation evidence for data and post-training choices because everything is released.
    Ai2 blog: Olmo 32 OLMo 2 Furious (2025-01)
  • 2025
    ★The Smol Training Playbook: The Secrets to Building World-Class LLMs
    Hugging Face (SmolLM team) · Oct 2025 · Book · Advanced · ~8 h
    A frank, practitioner-level account of training SmolLM3: ablation design, architecture choices, data mixtures, infrastructure, failures and mid-run fixes, then post-training.
    L7Lets you talk about how real pretraining decisions are made (ablation budgets, when to stop, what broke) rather than textbook recipes.
  • 2025
    On-Policy Distillation
    Thinking Machines Lab (Kevin Lu et al.) · Oct 2025 · Blog post · Advanced · ~1.5 h
    Explains why distilling on the student's own samples with dense teacher log-prob rewards beats off-policy SFT on teacher outputs, and how it compares to RL in compute.
    L7Lets you argue when to distill vs run RL for a smaller model, with cost reasoning.
    Hinton et al.: Distilling the Knowledge in a Neural Network (2015)
  • 2025
    ★nanochat
    Andrej Karpathy · Oct 2025 · Practice · Core · ~12 h
    Single small codebase that trains tokenizer, pretrains, mid-trains, SFTs and RLs a chat model for about $100 on one 8×H100 node (speedrun.sh).
    L7Gives you first-hand numbers for every stage of the pipeline in a weekend, so your answers carry real experience.
    Needs rented GPUs (~4 h on 8×H100) for the full speedrun; smaller configs run on one GPU.
    Walkthrough: Introducing nanochat (Discussion #1)
  • 2025
    LoRA Without Regret
    Thinking Machines Lab (John Schulman et al.) · Sep 2025 · Blog post · Advanced · ~1.5 h
    Careful experiments on when LoRA matches full fine-tuning (RL needs very little capacity; SFT on big data does not) and how to set its learning rate.
    L7Lets you size parameter-efficient post-training runs and explain why RL updates carry few bits.
  • 2025
    gpt-oss-120b & gpt-oss-20b Model Card
    OpenAI · Aug 2025 · Paper · Advanced · ~2 h
    OpenAI's open-weight MoE models: alternating dense and sliding-window attention, attention sinks, MXFP4 weights, harmony format and adjustable reasoning effort.
    L7The only recent first-party OpenAI architecture disclosure; useful for an OpenAI-specific critique.
  • 2025
    Post-training of LLMs
    Banghua Zhu; DeepLearning.AI · Jul 2025 · Course · Intro · ~1.5 h · video
    Nine short lessons with code contrasting SFT, DPO and online RL, and when each is the right tool.
    L7Baseline; a 90-minute refresher before reading the RLHF book.
    Free with a DeepLearning.AI login.
    Reinforcement Fine-Tuning LLMs With GRPO (Predibase, 2025-05)How Transformer LLMs Work (Alammar & Grootendorst)
  • 2025
    ★The Big LLM Architecture Comparison
    Sebastian Raschka (Ahead of AI) · Jul 2025 · Blog post · Core · ~3 h
    Side-by-side diagrams of DeepSeek-V3, OLMo 2, Gemma 3, Llama 4, Qwen3, Kimi K2, gpt-oss and others: MLA vs GQA, norm placement, QK-norm, MoE layout, sliding windows.
    L7The fastest way to build the curriculum's one-page architecture diff table and say why each lab chose what it did.
    Updated after first publication as new models came out.
    LLM Architecture Gallery (kept up to date, incl. Qwen3-Next/3.5 hybrids)From GPT-2 to gpt-oss
  • 2025
    Kimi K2: Open Agentic Intelligence
    Kimi Team (Moonshot AI) · Jul 2025 · Paper · Advanced · ~4 h
    1T-parameter MoE trained on 15.5T tokens with the MuonClip optimizer (QK-clip to stop attention-logit blow-up), plus large-scale agentic data synthesis and RL.
    L7Lets you discuss optimizer choice and logit-growth instabilities at trillion-parameter scale with a concrete fix.
    Muon is Scalable for LLM Training (Moonshot, 2025)Keller Jordan: Muon blog post
  • 2025
    Why We Think
    Lilian Weng · May 2025 · Blog post · Advanced · ~3 h
    Long, careful review of test-time compute: chain-of-thought, parallel vs sequential sampling, RL for reasoning, latent thinking and whether CoT is faithful.
    L7Lets you discuss test-time compute as a resource-allocation problem and critique CoT faithfulness, beyond 'more tokens help'.
  • 2025
    Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free
    Zihan Qiu et al. (Qwen team) · May 2025 · Paper · Advanced · ~1.5 h
    Large ablation showing an output gate after attention removes attention sinks and improves stability; adopted in Qwen3-Next.
    L7A good example of a recent, well-ablated architecture change you can critique and connect to stability and long context.
    Efficient Streaming LMs with Attention Sinks (2023)Kimi Linear: hybrid linear attention (2025)
  • 2025
    Qwen3 Technical Report
    Qwen Team (Alibaba) · May 2025 · Paper · Advanced · ~4 h
    Dense and MoE family with QK-norm, no shared experts, 36T-token pretraining in three stages, and a single model with thinking and non-thinking modes plus strong-to-weak distillation.
    L7Lets you contrast Qwen's and DeepSeek's MoE choices and discuss thinking-budget control as a product decision.
  • 2025
    The Entropy Mechanism of Reinforcement Learning for Reasoning Language Models
    Ganqu Cui et al. (Shanghai AI Lab, Tsinghua and others) · May 2025 · Paper · Advanced · ~2 h
    Explains why policy entropy collapses early in RLVR, links it to the covariance of log-prob and advantage, and proposes Clip-Cov/KL-Cov fixes.
    L7Lets you answer the 'entropy collapse in RL' debugging scenario with a mechanism and a fix, not just 'raise the temperature'.
  • 2025
    Vision Language Models (Better, faster, stronger)
    Hugging Face · May 2025 · Blog post · Core · ~1.5 h
    Current survey-style overview of VLMs: any-to-any models, MoE decoders, small VLMs, video understanding, VLAs and alignment.
    L7Good map of how images and video get into an LLM today; skim, then go deep on one technical report.
    Vision Language Models Explained (2024)Flamingo: cross-attention VLM (DeepMind, 2022)Chameleon: early-fusion native image tokens (Meta, 2024)
  • 2025
    The State of Reinforcement Learning for LLM Reasoning
    Sebastian Raschka (Ahead of AI) · Apr 2025 · Blog post · Core · ~2 h
    Readable walk from PPO to GRPO with a digest of the 2025 papers that tweaked it (length bias, KL, data difficulty).
    L7Gives you a structured 'PPO vs GRPO in two minutes' answer with the follow-up papers ready.
    Understanding Reasoning LLMsThe State of LLMs 2025
  • 2025
    Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?
    Yang Yue et al. (Tsinghua) · Apr 2025 · Paper · Advanced · ~1.5 h
    Uses pass@k at large k to argue RLVR mostly sharpens sampling of reasoning the base model already has rather than adding new capability.
    L7A ready critique of reasoning-RL claims; pair with ProRL's counter-evidence to show you can weigh conflicting results.
    ProRL: Prolonged RL Expands Reasoning Boundaries (NVIDIA, 2025)
  • 2025
    The Leaderboard Illusion
    Shivalika Singh et al. (Cohere Labs and others) · Apr 2025 · Paper · Advanced · ~1.5 h
    Shows how private testing, selective disclosure and data access skew Chatbot Arena rankings.
    L7Lets you challenge benchmark-based claims about a lab's model, a key move in 'what's wrong with our models'.
  • 2025
    A Comprehensive Survey on Long Context Language Modeling
    Jiaheng Liu et al. · Mar 2025 · Survey paper · Core · ~4 h
    Positional encodings and extrapolation, context-extension training, efficient attention, KV-cache management and long-context evaluation in one place.
    L7Lets you answer 'different ways of encoding position' and 'how would you get to 1M context' with a full menu and trade-offs.
  • 2025
    A Survey on Test-Time Scaling in Large Language Models: What, How, Where, and How Well?
    Qiyuan Zhang et al. · Mar 2025 · Survey paper · Core · ~3 h
    Organizes test-time compute by what is scaled (parallel, sequential, hybrid, internal), how, where it is applied and how it is evaluated.
    L7Gives a framework for comparing best-of-N, search and long-CoT approaches by cost per accuracy point.
    From System 1 to System 2: A Survey of Reasoning LLMs
  • 2025
    DAPO: An Open-Source LLM Reinforcement Learning System at Scale
    Qiying Yu et al. (ByteDance Seed, Tsinghua) · Mar 2025 · Paper · Advanced · ~2 h
    Four GRPO fixes for long-CoT RL: decoupled clip ranges against entropy collapse, dynamic sampling, token-level loss and overlong-reward shaping.
    L7Lets you discuss failure modes of GRPO at scale (entropy collapse, length bias) and the specific fixes.
    Understanding R1-Zero-Like Training: A Critical Perspective (Dr. GRPO)VinePPO: credit assignment in LLM RL
  • 2025
    Deep Dive into LLMs like ChatGPT
    Andrej Karpathy · Feb 2025 · Lecture · Intro · ~3.5 h · video
    Plain-language tour of the full pipeline: data, tokenization, pretraining, SFT, RLHF and RL for reasoning, with good intuition for hallucination and why models fail.
    L7Baseline; the fastest single overview if the post-training stack is new to you.
  • 2025
    How To Scale Your Model
    Google DeepMind (Jacob Austin et al.) · Feb 2025 · Book · Advanced · ~8 h
    Roofline analysis, FLOP and memory accounting for Transformers, and how to choose parallelism; the clearest source for doing scaling arithmetic on paper.
    L7Lets you estimate a 70B run (FLOPs, GPU-days, memory per device) live in an interview, with numbers.
    Distributed-training chapters overlap another track; read the Transformer-math and roofline chapters here.
  • 2025
    LLM Post-Training: A Deep Dive into Reasoning Large Language Models
    Komal Kumar, Tajamul Ashraf et al. (MBZUAI and others) · Feb 2025 · Survey paper · Core · ~4 h
    Survey of fine-tuning, RL (PPO, DPO, GRPO and variants) and test-time scaling methods for reasoning models.
    L7Reference map to check coverage of post-training methods after the RLHF book.
  • 2025
    Qwen2.5-VL Technical Report
    Qwen Team (Alibaba) · Feb 2025 · Paper · Advanced · ~3 h
    Native-resolution ViT with window attention, dynamic frame-rate sampling and absolute-time MRoPE for long video.
    L7Lets you explain concretely what changes from images to video (token budget, temporal position encoding, frame sampling).
  • 2025
    The Illustrated DeepSeek-R1
    Jay Alammar · Jan 2025 · Blog post · Intro · ~1 h
    Diagram-first explanation of the R1 pipeline: R1-Zero pure RL, cold-start SFT, reasoning RL, rejection sampling and distillation.
    L7Baseline; read before the R1 paper to keep the stages straight.
    The Illustrated GPT-OSS
  • 2025
    AI Engineering (Ch. 7 Finetuning, Ch. 8 Dataset Engineering)
    Chip Huyen (O'Reilly) · Jan 2025 · Book · Intro · ~5 h · paid
    Practitioner view of when to fine-tune vs prompt or retrieve, memory math for fine-tuning, PEFT, model merging and building fine-tuning data.
    L7Baseline; useful for the applied framing, not for research depth.
    Paid. O'Reilly page returns 403 to scripts; author's book page verified.
    Book resources (GitHub)
  • 2025
    ★DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
    DeepSeek-AI · Jan 2025 · Paper · Advanced · ~3 h
    Shows pure RL with rule-based rewards produces long chain-of-thought (R1-Zero), then the multi-stage R1 recipe and distillation into small dense models.
    L7Lets you explain what RLVR does and does not need (no process reward model, no MCTS) and why distillation beats RL for small models.
  • 2025
    Hugging Face LLM Course
    Hugging Face · 2025 · Course · Intro · ~15 h
    Hands-on notebooks for tokenizers, fine-tuning and, in Chapter 12, building a reasoning model with GRPO; quickest way to get working TRL code.
    L7Baseline; use it for working code, not for interview depth.
    Chapter 12: Open R1 for Students (GRPO)Hugging Face Deep RL Course (PPO unit)
  • 2024
    ★DeepSeek-V3 Technical Report
    DeepSeek-AI · Dec 2024 · Paper · Advanced · ~6 h
    MLA, fine-grained MoE with shared experts, auxiliary-loss-free load balancing, multi-token prediction, FP8 training and DualPipe, with costs.
    L7Lets you discuss how architecture, numerics and systems were co-designed for a fixed GPU budget, the deepest question in this track.
    DeepSeekMoE (2024)DeepSeek-V3.2: DeepSeek Sparse Attention (2025-12)
  • 2024
    Reward Hacking in Reinforcement Learning
    Lilian Weng · Nov 2024 · Blog post · Advanced · ~2 h
    Survey-style post on reward hacking in RL and RLHF: over-optimization, sycophancy, test-tampering in coding RL, and mitigations.
    L7Lets you name the likely failure mode in a lab's post-trained model and propose a measurable fix.
    Policy Gradient Algorithms (Weng, background)
  • 2024
    Tulu 3: Pushing Frontiers in Open Language Model Post-Training
    Nathan Lambert et al. (Ai2) · Nov 2024 · Paper · Advanced · ~4 h
    Fully documented modern post-training: prompt curation and decontamination, SFT mixes, on-policy DPO, and RL with verifiable rewards (RLVR), with an evaluation suite.
    L7The best open recipe to cite when asked to design a post-training pipeline from scratch.
  • 2024
    Asynchronous RLHF: Faster and More Efficient Off-Policy RL for Language Models
    Michael Noukhovitch et al. (Mila, Ai2) · Oct 2024 · Paper · Advanced · ~2 h
    Separates generation and learning so they run concurrently, and measures how much off-policy staleness online DPO/PPO tolerate.
    L7Gives you the core trade-off (throughput vs staleness) for designing an async RL post-training system, as asked in a Meta MSL round.
    AReaL: fully asynchronous RL system for reasoning (2025)
  • 2024
    MIT 6.5940: TinyML and Efficient Deep Learning Computing (Fall 2024)
    Song Han (MIT HAN Lab) · Sep 2024 · Course · Core · ~20 h · video
    Efficiency lectures on quantization, pruning, distillation, efficient attention and LLM serving; explains why GQA/MLA/MoE exist from the memory-bandwidth side.
    L7Lets you tie architecture choices to inference cost (KV-cache size, bandwidth, quantization error) in numbers.
    Fall 2024 is the latest offering with a full public video set; Fall 2026 is in progress.
    Course hub (Fall 2026 offering)MIT HAN Lab YouTube
  • 2024
    COS 597R: Deep Dive into Large Language Models (Fall 2024)
    Danqi Chen, Sanjeev Arora (Princeton) · Sep 2024 · Course · Advanced · ~15 h
    Research-seminar reading list with lecture slides on pretraining, data, scaling, post-training and reasoning, chosen by two leading researchers.
    L7Use the reading list as a checklist of papers a research-track interviewer may expect you to have opinions on.
    No videos; slides and readings only.
    COS 597G (Fall 2022): Understanding LLMs
  • 2024
    Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters
    Charlie Snell, Jaehoon Lee, Kelvin Xu, Aviral Kumar (Berkeley, Google DeepMind) · Aug 2024 · Paper · Advanced · ~2 h
    Compares best-of-N with verifiers, beam search with process rewards, and sequential revision, and shows the best strategy depends on problem difficulty.
    L7Lets you argue test-time vs pretraining compute trade-offs with a compute-optimal framing.
    s1: Simple test-time scaling (budget forcing, 2025)
  • 2024
    A Survey on Mixture of Experts in Large Language Models
    Weilin Cai, Juyong Jiang, Fan Wang, Jing Tang, Sunghun Kim, Jiayi Huang · Jul 2024 · Survey paper · Core · ~4 h
    Covers routing, load balancing, expert granularity, shared experts and system design for MoE LLMs.
    L7Gives the vocabulary for the Mistral-style 'MoE and expert parallelism' question; pair with DeepSeek-V3 for the current recipe.
  • 2024
    ★The Llama 3 Herd of Models
    Llama Team, AI @ Meta · Jul 2024 · Paper · Advanced · ~8 h
    The most complete public account of a frontier dense-model pipeline: data mix and annealing, scaling-law-chosen size, 4D parallelism, reliability at 16K GPUs, and iterated SFT + rejection sampling + DPO.
    L7The curriculum's bar is to explain every design choice here; doing so shows you can reason about a full frontier pipeline.
  • 2024
    Let's reproduce GPT-2 (124M)
    Andrej Karpathy · Jun 2024 · Lecture · Core · ~4 h · video
    Four hours from empty file to a trained GPT-2, including mixed precision, FlashAttention, learning-rate schedule, gradient accumulation and DDP.
    L7Lets you talk through the real training-loop details (warmup, weight decay groups, loss spikes) the way someone who has run one does.
    Let's build GPT: from scratch (2023)nanoGPT repo
  • 2024
    The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale
    Guilherme Penedo et al. (Hugging Face) · Jun 2024 · Paper · Advanced · ~3 h
    Detailed ablations of every web-data step (extraction, filtering, dedup granularity) and the LLM-annotated educational-quality classifier behind FineWeb-Edu.
    L7Lets you design and defend a pretraining data pipeline with ablation evidence, a common open-ended interview prompt.
    FineWeb blog post (interactive)DataComp-LM (DCLM, 2024)
  • 2024
    DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model
    DeepSeek-AI · May 2024 · Paper · Advanced · ~3 h
    Introduces Multi-head Latent Attention (low-rank KV compression with decoupled RoPE); the clearest derivation of MLA.
    L7Lets you compare MLA vs GQA on KV-cache bytes per token and explain why decoupled RoPE is needed.
  • 2024
    LoRA Learns Less and Forgets Less
    Dan Biderman et al. (Columbia, Databricks) · May 2024 · Paper · Advanced · ~1.5 h
    Measures the learning-vs-forgetting trade-off of LoRA vs full fine-tuning on code and math, with continued pretraining and instruction tuning.
    L7Lets you discuss catastrophic forgetting and domain adaptation with evidence (rank, data volume, replay) rather than folklore.
    Don't Stop Pretraining: domain- and task-adaptive pretraining (2020)
  • 2024
    Automatic Data Curation for Self-Supervised Learning: A Clustering-Based Approach
    Huy V. Vo et al. (Meta FAIR) · May 2024 · Paper · Advanced · ~1.5 h
    Hierarchical k-means over embeddings to build balanced, diverse datasets from huge unlabeled image pools without labels.
    L7Gives an embedding-and-clustering backbone for 'mine novel images from a huge unlabeled corpus', with labelers spent only on cluster samples.
    A Survey of Deep Active Learning (2020)
  • 2024
    Let's build the GPT Tokenizer
    Andrej Karpathy · Feb 2024 · Lecture · Core · ~2 h · video
    Builds byte-level BPE from scratch and shows the failure modes (arithmetic, trailing spaces, glitch tokens, multilingual cost) that trace back to tokenization.
    L7Lets you explain model weaknesses that come from the tokenizer, a common thread in 'what's wrong with our model' questions.
  • 2024
    A Survey on Data Selection for Language Models
    Alon Albalak et al. · Feb 2024 · Survey paper · Core · ~4 h
    Taxonomy of data selection for pretraining, instruction tuning and preference data: heuristic filters, quality classifiers, dedup, domain mixing.
    L7Lets you design a data pipeline end to end and say which filter you'd ablate first.
  • 2024
    ★DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models
    Zhihong Shao et al. (DeepSeek-AI) · Feb 2024 · Paper · Advanced · ~2 h
    Introduces GRPO (group-relative advantages, no value network) and a unified view of SFT, RFT, DPO, PPO and GRPO as gradient-coefficient choices.
    L7Source for the 'PPO vs GRPO' answer; its unified-paradigm section lets you compare methods on one axis.
  • 2024
    Mixtral of Experts
    Mistral AI · Jan 2024 · Paper · Core · ~1 h
    Sparse MoE with 8 experts and top-2 routing: ~47B total parameters but ~13B active per token, with routing analysis.
    L7Directly answers the Mistral 'why does 47B run like 13B' question; know where the memory cost still lands.
  • 2023
    Beyond Chinchilla-Optimal: Accounting for Inference in Language Model Scaling Laws
    Nikhil Sardana et al. (MosaicML/Databricks) · Dec 2023 · Paper · Advanced · ~1.5 h
    Adds inference demand to the scaling-law objective and shows why labs train smaller models on far more tokens than Chinchilla-optimal.
    L7Lets you answer 'why do labs over-train' with a total-cost argument instead of a slogan.
  • 2023
    Mamba: Linear-Time Sequence Modeling with Selective State Spaces
    Albert Gu, Tri Dao · Dec 2023 · Paper · Advanced · ~2.5 h
    The leading attention-free architecture: input-dependent state-space layers with a hardware-aware scan, linear in sequence length.
    L7Lets you argue the trade-offs of attention-free and hybrid models (recall, in-context retrieval, inference cost), which now appear in Qwen3-Next and Kimi Linear.
  • 2023
    The N Implementation Details of RLHF with PPO
    Shengyi Huang et al. (Hugging Face) · Oct 2023 · Blog post · Advanced · ~2 h
    Reproduces OpenAI's original RLHF code and lists the details that matter: reward normalization, KL penalty handling, padding, value-head init, whitening.
    L7Lets you answer 'your PPO run diverged, what do you check' with specific knobs, a staff-level debugging signal.
    The 37 Implementation Details of PPO (ICLR blog 2022)Illustrating RLHF (HF blog)
  • 2023
    Mistral 7B
    Mistral AI · Oct 2023 · Paper · Core · ~1 h
    Short report combining grouped-query attention with sliding-window attention and a rolling KV cache.
    L7Lets you say what GQA and SWA each buy (KV memory, long-sequence cost) and what they give up (heads, direct long-range attention).
  • 2023
    YaRN: Efficient Context Window Extension of Large Language Models
    Bowen Peng, Jeffrey Quesnelle, Honglu Fan, Enrico Shippole · Sep 2023 · Paper · Advanced · ~1.5 h
    The RoPE-scaling method (frequency interpolation plus attention temperature) used by Qwen, DeepSeek and others to extend context after pretraining.
    L7Lets you explain how labs go from 4K to 128K+ context cheaply and what breaks at long range.
    Round and Round We Go! What makes RoPE useful? (2024)
  • 2023
    Small-scale proxies for large-scale Transformer training instabilities
    Mitchell Wortsman et al. (Google DeepMind) · Sep 2023 · Paper · Advanced · ~2 h
    Reproduces large-run instabilities (attention-logit growth, output-logit divergence) in small models and tests fixes: QK-layernorm, z-loss, warmup, weight decay, μParam.
    L7Lets you diagnose a loss spike and propose a cheap small-scale experiment before touching the big run.
  • 2023
    Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback
    Stephen Casper et al. · Jul 2023 · Survey paper · Advanced · ~3 h
    Structured list of what goes wrong in RLHF: human feedback quality, reward-model misspecification, policy over-optimization, and which problems are fundamental.
    L7Lets you critique any lab's alignment pipeline systematically rather than with one-off anecdotes.
  • 2023
    Llama 2: Open Foundation and Fine-Tuned Chat Models
    Hugo Touvron et al. (Meta) · Jul 2023 · Paper · Advanced · ~4 h
    The most detailed public account of a human-preference collection pipeline: annotation protocol, separate helpfulness and safety reward models, iterative rejection sampling + PPO.
    L7Lets you design an RLHF feedback pipeline with real scale numbers (1.4M comparisons) and the reasons for each design choice.
  • 2023
    Lost in the Middle: How Language Models Use Long Contexts
    Nelson F. Liu et al. (Stanford) · Jul 2023 · Paper · Core · ~1 h
    Shows accuracy drops when relevant information sits mid-context, i.e. a longer window is not the same as usable context.
    L7Baseline evidence for 'what happens when you grow the context window'; pair with the long-context survey for fixes.
  • 2023
    GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints
    Joshua Ainslie et al. (Google) · May 2023 · Paper · Core · ~1 h
    Short paper introducing grouped-query attention, the KV-cache saving now used in Llama, Qwen and most dense models.
    L7Lets you quantify KV-cache savings of MHA vs MQA vs GQA and the quality cost.
  • 2023
    Direct Preference Optimization: Your Language Model is Secretly a Reward Model
    Rafael Rafailov et al. (Stanford) · May 2023 · Paper · Core · ~2 h
    Derives a closed-form loss that optimizes the RLHF objective directly from preference pairs without a reward model or sampling.
    L7Lets you derive DPO from the KL-regularized objective and argue offline-vs-online trade-offs (the Mistral 'RM-RLHF vs DPO' question).
  • 2023
    Visual Instruction Tuning (LLaVA)
    Haotian Liu, Chunyuan Li, Qingyang Wu, Yong Jae Lee · Apr 2023 · Paper · Core · ~1.5 h
    The simplest working vision-language recipe: frozen vision encoder, projector, LLM, and GPT-generated visual instruction data.
    L7Baseline for multimodal questions; know the encoder-projector-LLM pattern and its limits.
    Molmo and PixMo: open VLM data (Ai2, 2024)
  • 2022
    Constitutional AI: Harmlessness from AI Feedback
    Yuntao Bai et al. (Anthropic) · Dec 2022 · Paper · Core · ~2 h
    RL from AI feedback guided by a written constitution: self-critique SFT then RLAIF, reducing reliance on human harm labels.
    L7Essential context for Anthropic loops; lets you discuss scalable oversight and AI-feedback trade-offs.
  • 2022
    OPT-175B training chronicles (logbook)
    Meta AI (Susan Zhang et al.) · May 2022 · Reference · Advanced · ~2 h
    Day-by-day logbook of a 175B run: loss spikes, divergences, hardware failures, LR and clipping changes and restarts from checkpoints.
    L7Rare first-hand record of how a real team debugged divergence at scale; lets you narrate a credible debugging sequence.
    OPT paper (2022)
  • 2022
    Efficient Transformers: A Survey
    Yi Tay, Mostafa Dehghani, Dara Bahri, Donald Metzler (Google) · Mar 2022 · Survey paper · Core · ~3 h
    Taxonomy of sub-quadratic attention (sparse, low-rank, kernel, recurrence); older but still the standard map of the design space.
    L7Lets you place new hybrids (linear attention, sparse attention) in a known taxonomy and say what each gives up.
    v1 2020-09, last revised 2022-03; predates linear-attention hybrids (see r45).
  • 2022
    ★Training Compute-Optimal Large Language Models (Chinchilla)
    Jordan Hoffmann et al. (DeepMind) · Mar 2022 · Paper · Core · ~2 h
    Three methods for fitting compute-optimal params vs tokens (~20 tokens/param) and the correction to Kaplan's scaling laws.
    L7Lets you do C ≈ 6ND arithmetic and explain where the 20:1 rule comes from and when it does not apply.
    Kaplan et al.: Scaling Laws for Neural Language Models (2020)Chinchilla Scaling: A replication attempt (Epoch AI, 2024)
  • 2022
    Training language models to follow instructions with human feedback (InstructGPT)
    Long Ouyang et al. (OpenAI) · Mar 2022 · Paper · Core · ~2 h
    The canonical SFT → reward model → PPO recipe with the KL penalty and pretraining mix, plus the labeler process.
    L7Baseline for every RLHF question; know the three stages and the alignment tax result cold.
  • 2022
    Mixture-of-Experts with Expert Choice Routing
    Yanqi Zhou et al. (Google) · Feb 2022 · Paper · Advanced · ~1.5 h
    Experts pick tokens instead of tokens picking experts, giving perfect load balance and variable compute per token.
    L7Lets you compare token-choice top-k vs expert-choice routing, including why expert choice is awkward for autoregressive decoding.
    GShard: conditional computation and expert sharding (2020)
  • 2021
    Between words and characters: A Brief History of Open-Vocabulary Modeling and Tokenization in NLP
    Sabrina J. Mielke et al. · Dec 2021 · Survey paper · Core · ~2 h
    Canonical tokenization survey: word, subword (BPE, WordPiece, unigram), byte and character models, and their trade-offs.
    L7Baseline; enough to compare BPE vs unigram vs byte-level and argue vocabulary-size trade-offs.
    Older but still the standard survey; for tokenizer-free work see the Byte Latent Transformer link.
    Sennrich et al.: Neural Machine Translation of Rare Words with Subword Units (BPE, 2015)Byte Latent Transformer (Meta, 2024)
  • 2021
    Deduplicating Training Data Makes Language Models Better
    Katherine Lee et al. (Google) · Jul 2021 · Paper · Core · ~1.5 h
    Exact-substring (suffix array) and near-duplicate (MinHash) dedup at web scale, with effects on memorization and eval contamination.
    L7Gives the two dedup algorithms you'd combine in a streaming pipeline; pair with datatrove for the implementation.
    datatrove: Hugging Face's data-processing library (MinHash dedup, filters)
  • 2021
    RoFormer: Enhanced Transformer with Rotary Position Embedding
    Jianlin Su et al. · Apr 2021 · Paper · Core · ~1.5 h
    Introduces RoPE, the default positional scheme in nearly every current LLM; work through the rotation-matrix derivation.
    L7Lets you derive why RoPE gives relative position through a dot product and why it decays with distance.
    ALiBi: Train Short, Test Long (2021)
  • 2021
    Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity
    William Fedus, Barret Zoph, Noam Shazeer (Google) · Jan 2021 · Paper · Core · ~2 h
    Top-1 routing, capacity factor, auxiliary load-balancing loss and the training-stability tricks that made sparse MoE work.
    L7Lets you explain where MoE instability and load imbalance come from before discussing DeepSeek's fixes.
    Shazeer et al.: Sparsely-Gated Mixture-of-Experts (2017)
  • 2019
    Root Mean Square Layer Normalization
    Biao Zhang, Rico Sennrich · Oct 2019 · Paper · Core · ~1 h
    Short paper behind RMSNorm: drops mean-centring from LayerNorm for speed with little quality loss.
    L7Answers the 'RMSNorm vs LayerNorm' question directly; pair with pre-norm vs post-norm and QK-norm from r18.
    Shazeer: GLU Variants Improve Transformer (SwiGLU, 2020)
  • 2019
    The Curious Case of Neural Text Degeneration
    Ari Holtzman et al. (UW, Ai2) · Apr 2019 · Paper · Core · ~1.5 h
    Shows why greedy and beam search degenerate into repetition and introduces nucleus (top-p) sampling; the reference for comparing decoding strategies.
    L7Baseline; enough to explain when each decoding method fails and how temperature changes the distribution.
    Hugging Face: How to generate text (decoding methods)
  • 2019
    Language Models are Unsupervised Multitask Learners (GPT-2)
    Alec Radford et al. (OpenAI) · Feb 2019 · Paper · Core · ~1.5 h
    The decoder-only, pre-norm baseline every architecture tour starts from; also the WebText data idea.
    L7Baseline; the 'before' column of your architecture diff table.
    GPT-3: Language Models are Few-Shot Learners (2020)
  • 2018
    Spinning Up in Deep RL
    Joshua Achiam (OpenAI) · Nov 2018 · Practice · Core · ~8 h
    Compact RL theory notes plus clean reference implementations of VPG, PPO and others to code against.
    L7Practice target for a 'code an RL algorithm' screen: write REINFORCE and PPO-clip from memory, then port to an LLM token-level loss.
    Older but still the standard compact RL implementation reference.
    PPO page
  • 2017
    Proximal Policy Optimization Algorithms
    John Schulman et al. (OpenAI) · Jul 2017 · Paper · Core · ~1 h
    Short paper defining the clipped surrogate objective and GAE-based training loop that LLM RLHF inherited.
    L7Needed to explain exactly what GRPO removes (the value function) and keeps (clipping, KL).
  • 2017
    Attention Is All You Need
    Ashish Vaswani et al. (Google) · Jun 2017 · Paper · Core · ~2 h
    The original Transformer; read to know what has changed since (post-norm, sinusoidal positions, ReLU FFN, encoder-decoder).
    L7Baseline; skim if known, but be able to list every component modern models replaced and why.
Not found or not verified. CMU 11-667 (LLMs: Methods and Applications): cmu-llms.org now hosts 11-766 LLM Applications (Spring 2026), and I found no current 11-667 term page, so it is left out. Stanford CS324 (Winter 2022) exists but is superseded by CS336, so it is left out. Public 2026 videos are not confirmed for CS224N, CS234, CS285 or CMU 11-711, so older public playlists are linked instead. MIT 6.5940 Fall 2026 is running, but Fall 2024 is the latest offering with a full video set. The DeepSeek-V4 report is linked from its Hugging Face card; I found no arXiv version. Qwen3-Next/Qwen3.5 have no official technical report I could find, so the Raschka gallery covers them. OpenAI's sycophancy postmortems (openai.com) return 403 to scripts and were not verified, so they are omitted. I found no strong 2025–26 tokenization survey to replace Mielke et al. (2021). The Hugging Face Ultra-Scale Playbook is left to the distributed-training track. All 48 collected question groups are covered (coverage array); none uncovered. No item in the brief's recurring-problem list applies, so every `serves` is []. Links checked 27 Oct 2026.
← PreviousML coding from scratch and debuggingNext →Project deep dive and leadership stories
01 Practical coding02 Algorithms03 Fundamentals04 ML coding05 LLM internals06 Deep dive & leadership07 System design08 Inference design09 ML infra10 Training at scale11 Applied LLM12 Research13 Values & mission