Resources & reading

A curated, opinionated reading list across the pipeline. Not exhaustive — meant to be read.

For inspiration, see Andrej Karpathy's original knowledge-base gist and his Intro to LLMs talk.

Foundational papers (must-read)

• Vaswani et al., "Attention Is All You Need", 2017. The Transformer. arxiv.org/abs/1706.03762
• Brown et al., "Language Models are Few-Shot Learners" (GPT-3), 2020. arxiv.org/abs/2005.14165
• Kaplan et al., "Scaling Laws", 2020. arxiv.org/abs/2001.08361
• Hoffmann et al., "Chinchilla", 2022. arxiv.org/abs/2203.15556
• Ouyang et al., "InstructGPT", 2022. arxiv.org/abs/2203.02155
• Bai et al., "Constitutional AI", 2022. arxiv.org/abs/2212.08073
• Rafailov et al., "DPO", 2023. arxiv.org/abs/2305.18290

Reference modern recipes (open-weights)

• Llama 3 paper, Meta, 2024. The most comprehensive single-paper modern recipe. arxiv.org/abs/2407.21783
• OLMo 2, AI2, 2024. Fully open weights, data, recipe. arxiv.org/abs/2501.00656
• Tülu 3, AI2, 2024. The clearest open post-training recipe with RLVR. arxiv.org/abs/2411.15124
• DeepSeek-V3 technical report, 2024. SOTA open MoE. arxiv.org/abs/2412.19437
• DeepSeek-R1, 2025. RL-trained reasoning model. arxiv.org/abs/2501.12948
• Qwen 2.5 technical report, 2024. Strong open recipe with detailed report.
• MiniCPM technical report, 2024. Excellent practical writeup of small-model training. arxiv.org/abs/2404.06395
• Mixtral 8×7B paper, Mistral, 2024. Reference open MoE. arxiv.org/abs/2401.04088

Data

• The Pile — Gao et al., 2020. arxiv.org/abs/2101.00027
• RefinedWeb — Penedo et al., 2023. arxiv.org/abs/2306.01116
• FineWeb / FineWeb-Edu — HuggingFace, 2024. huggingface.co/spaces/HuggingFaceFW/blogpost-fineweb-v1
• Dolma — AI2, 2024. arxiv.org/abs/2402.00159
• DataComp-LM, 2024. arxiv.org/abs/2406.11794
• Lee et al., "Deduplicating Training Data", 2021. arxiv.org/abs/2107.06499
• Carlini et al., "Quantifying Memorization", 2022. arxiv.org/abs/2202.07646
• Albalak et al., "A Survey on Data Selection", 2024. arxiv.org/abs/2402.16827
• CulturaX — multilingual web. arxiv.org/abs/2309.09400
• The Stack v2 — code corpus. huggingface.co/datasets/bigcode/the-stack-v2

Architecture

• RoPE — Su et al., 2021. arxiv.org/abs/2104.09864
• GQA — Ainslie et al., 2023. arxiv.org/abs/2305.13245
• SwiGLU — Shazeer, 2020. arxiv.org/abs/2002.05202
• RMSNorm — Zhang & Sennrich, 2019. arxiv.org/abs/1910.07467
• The Annotated Transformer — Harvard NLP. nlp.seas.harvard.edu/annotated-transformer

Pretraining mechanics

• µP (Tensor Programs V) — Yang et al., 2022. arxiv.org/abs/2203.03466
• ZeRO — Rajbhandari et al., 2019. arxiv.org/abs/1910.02054
• Megatron-LM — Shoeybi et al., 2019. arxiv.org/abs/1909.08053
• FlashAttention — Dao et al., 2022. arxiv.org/abs/2205.14135
• FlashAttention-2 — Dao, 2023. arxiv.org/abs/2307.08691

Long context

• Position Interpolation — Chen et al., 2023. arxiv.org/abs/2306.15595
• YaRN — Peng et al., 2023. arxiv.org/abs/2309.00071
• LongRoPE — Microsoft, 2024. arxiv.org/abs/2402.13753
• Lost in the Middle — Liu et al., 2023. arxiv.org/abs/2307.03172
• RULER — Hsieh et al., 2024. arxiv.org/abs/2404.06654

Post-training

• InstructGPT — Ouyang et al., 2022. arxiv.org/abs/2203.02155
• HHH RLHF — Bai et al., 2022. arxiv.org/abs/2204.05862
• PPO — Schulman et al., 2017. arxiv.org/abs/1707.06347
• DPO — Rafailov et al., 2023. arxiv.org/abs/2305.18290
• IPO — Azar et al., 2023. arxiv.org/abs/2310.12036
• KTO — Ethayarajh et al., 2024. arxiv.org/abs/2402.01306
• ORPO — Hong et al., 2024. arxiv.org/abs/2403.07691
• SimPO — Meng et al., 2024. arxiv.org/abs/2405.14734
• GRPO / DeepSeekMath — Shao et al., 2024. arxiv.org/abs/2402.03300
• RLVR / Tülu 3 — Lambert et al., 2024. arxiv.org/abs/2411.15124
• Self-Rewarding LMs — Yuan et al., 2024. arxiv.org/abs/2401.10020
• LIMA — Zhou et al., 2023. arxiv.org/abs/2305.11206
• UltraFeedback — Cui et al., 2023. arxiv.org/abs/2310.01377
• Length bias in RLHF — Singhal et al., 2023. arxiv.org/abs/2310.03716
• RewardBench — Lambert et al., 2024. arxiv.org/abs/2403.13787
• Lightman et al., "Let's Verify Step by Step", 2023. arxiv.org/abs/2305.20050

Reasoning

• Wei et al., "Chain-of-Thought", 2022. arxiv.org/abs/2201.11903
• GSM8K — Cobbe et al., 2021. arxiv.org/abs/2110.14168
• MATH — Hendrycks et al., 2021. arxiv.org/abs/2103.03874
• DeepSeek-R1, 2025. arxiv.org/abs/2501.12948
• GPQA — Rein et al., 2023. arxiv.org/abs/2311.12022

Tool use & agents

• Toolformer — Schick et al., 2023. arxiv.org/abs/2302.04761
• ReAct — Yao et al., 2022. arxiv.org/abs/2210.03629
• Gorilla — Patil et al., 2023. arxiv.org/abs/2305.15334
• SWE-bench — Jimenez et al., 2024. arxiv.org/abs/2310.06770
• τ-bench, 2024. arxiv.org/abs/2406.12045
• WebArena, 2023. arxiv.org/abs/2307.13854
• Berkeley Function-Calling Leaderboard — gorilla.cs.berkeley.edu/leaderboard.html

RAG

• Lewis et al., "RAG", 2020. arxiv.org/abs/2005.11401
• Karpukhin et al., "DPR", 2020. arxiv.org/abs/2004.04906
• Self-RAG — Asai et al., 2023. arxiv.org/abs/2310.11511
• RAG survey — Gao et al., 2023. arxiv.org/abs/2312.10997

Structured output

• Outlines paper — Willard & Louf, 2023. arxiv.org/abs/2307.09702
• XGrammar — Dong et al., 2024. arxiv.org/abs/2411.15100
• OpenAI Structured Outputs — platform.openai.com/docs/guides/structured-outputs

Safety & red teaming

• Constitutional AI — Bai et al., 2022. arxiv.org/abs/2212.08073
• Llama Guard — Inan et al., 2023. arxiv.org/abs/2312.06674
• WildGuard — Han et al., 2024. arxiv.org/abs/2406.18495
• XSTest — Röttger et al., 2023. arxiv.org/abs/2308.01263
• HarmBench — Mazeika et al., 2024. arxiv.org/abs/2402.04249
• GCG attacks — Zou et al., 2023. arxiv.org/abs/2307.15043
• Indirect prompt injection — Greshake et al., 2023. arxiv.org/abs/2302.12173
• Anthropic Acceptable Use Policy — anthropic.com/legal/aup
• OpenAI Model Spec — openai.com/index/openai-model-spec

Multimodal

• CLIP — Radford et al., 2021. arxiv.org/abs/2103.00020
• SigLIP — Zhai et al., 2023. arxiv.org/abs/2303.15343
• LLaVA — Liu et al., 2023. arxiv.org/abs/2304.08485
• Qwen-VL — Bai et al., 2023. arxiv.org/abs/2308.12966
• Idefics2 — Laurençon et al., 2024. arxiv.org/abs/2405.02246
• MMMU benchmark, 2023. arxiv.org/abs/2311.16502

Compression & efficiency

• GPTQ — Frantar et al., 2022. arxiv.org/abs/2210.17323
• AWQ — Lin et al., 2023. arxiv.org/abs/2306.00978
• LoRA — Hu et al., 2021. arxiv.org/abs/2106.09685
• QLoRA — Dettmers et al., 2023. arxiv.org/abs/2305.14314
• Speculative decoding — Leviathan et al., 2022. arxiv.org/abs/2211.17192
• vLLM / PagedAttention — Kwon et al., 2023. arxiv.org/abs/2309.06180

Evaluation

• HELM — Liang et al., 2022. arxiv.org/abs/2211.09110
• MMLU — Hendrycks et al., 2020. arxiv.org/abs/2009.03300
• BIG-Bench Hard — Suzgun et al., 2022. arxiv.org/abs/2210.09261
• MT-Bench / Arena — Zheng et al., 2023. arxiv.org/abs/2306.05685
• AlpacaEval 2 LC — Dubois et al., 2024. arxiv.org/abs/2404.04475
• IFEval — Zhou et al., 2023. arxiv.org/abs/2311.07911
• SimpleQA — OpenAI, 2024. arxiv.org/abs/2411.04368

Tooling and frameworks

Training

• nanoGPT — minimal Transformer training. github.com/karpathy/nanoGPT
• llm.c — Karpathy. CUDA Transformer training. github.com/karpathy/llm.c
• torchtitan — PyTorch reference distributed training. github.com/pytorch/torchtitan
• nanotron — HuggingFace's clean training library. github.com/huggingface/nanotron
• Megatron-LM / Megatron-Core — NVIDIA. github.com/NVIDIA/Megatron-LM
• DeepSpeed — Microsoft. github.com/microsoft/DeepSpeed

Tokenization

• minbpe — Karpathy. github.com/karpathy/minbpe
• HuggingFace tokenizers — github.com/huggingface/tokenizers
• tiktoken — OpenAI. github.com/openai/tiktoken
• sentencepiece — Google. github.com/google/sentencepiece

Data

• datatrove — HuggingFace. github.com/huggingface/datatrove
• text-dedup — github.com/ChenghaoMou/text-dedup
• presidio — Microsoft PII detection. github.com/microsoft/presidio

Post-training

• trl — HuggingFace. github.com/huggingface/trl
• axolotl — community SFT/DPO. github.com/axolotl-ai-cloud/axolotl
• open-instruct — AI2 / Tülu. github.com/allenai/open-instruct
• OpenRLHF — distributed RLHF. github.com/OpenRLHF/OpenRLHF
• verl — open RL training framework. github.com/volcengine/verl
• distilabel — synthetic data. github.com/argilla-io/distilabel

Evaluation

• lm-evaluation-harness — EleutherAI. github.com/EleutherAI/lm-evaluation-harness
• inspect_ai — UK AISI. github.com/UKGovernmentBEIS/inspect_ai
• evals — OpenAI. github.com/openai/evals
• lmms-eval — multimodal. github.com/EvolvingLMMs-Lab/lmms-eval

Serving

• vLLM — github.com/vllm-project/vllm
• SGLang — github.com/sgl-project/sglang
• TensorRT-LLM — github.com/NVIDIA/TensorRT-LLM
• llama.cpp — github.com/ggerganov/llama.cpp
• MLC LLM — github.com/mlc-ai/mlc-llm

Constrained decoding

• Outlines — github.com/dottxt-ai/outlines
• xgrammar — github.com/mlc-ai/xgrammar
• lm-format-enforcer — github.com/noamgat/lm-format-enforcer

Red teaming

• pyrit — Microsoft. github.com/Azure/PyRIT
• garak — github.com/leondz/garak

Excellent talks and writeups

• Andrej Karpathy — "Intro to Large Language Models" (1 hr). youtube.com/watch?v=zjkBMFhNj_g
• Karpathy — "Let's build GPT, from scratch" (2 hr). youtube.com/watch?v=kCc8FmEb1nY
• Karpathy — "Let's build the GPT Tokenizer" (2 hr). youtube.com/watch?v=zduSFxRajkE
• Karpathy — "Let's reproduce GPT-2 (124M)". youtube.com/watch?v=l8pRSuU81PU
• Karpathy's original LLM knowledge-base gist — the inspiration for this wiki. gist.github.com/karpathy/442a6bf555914893e9891c11519de94f
• Nathan Lambert's blog (Interconnects) — best practical post-training writing on the open web. interconnects.ai
• HuggingFace blog — many excellent practical posts on data, post-training, and serving.
• Sebastian Raschka — "Build a Large Language Model (From Scratch)" — the best book-length intro. github.com/rasbt/LLMs-from-scratch
• Lilian Weng's blog — careful, well-cited primers on most topics. lilianweng.github.io