LLM Training, Interactively — Series Hub

The pipeline

Click any stage to jump to its part.

All 15 parts

Part 1 of 15

What is a language model

Next-token prediction, autoregressive generation, and why OLMo is the running example for this series.

Part 2 of 15

Tokenization & embeddings

Real BPE training, multilingual token cost, and learned (not hand-placed) word vectors.

Part 3 of 15

The transformer, block by block

Attention, multi-head, RoPE, residual stream, norms, SwiGLU, parameter counting, and MoE.

Part 4 of 15

Building the pretraining dataset

Sourcing, filtering, deduplication, contamination, and mixing at trillion-token scale.

Part 5 of 15

Pretraining: the training loop

Loss, perplexity, learning-rate schedules, gradient clipping, and mid-training annealing.

Part 6 of 15

Scaling laws & training systems

Chinchilla compute-optimal scaling, memory accounting, and distributed-training parallelism.

Part 7 of 15

Supervised fine-tuning & PEFT

Chat templates, loss masking, and LoRA/QLoRA for parameter-efficient fine-tuning.

Part 8 of 15

Alignment: RLHF & DPO

Reward models, PPO's clipped objective, DPO's implicit reward, and the KL-reward frontier.

Part 9 of 15

Reasoning & RL with verifiable rewards

Chain-of-thought, RLVR, GRPO, and test-time compute scaling.

Part 10 of 15

Evaluation

How benchmarks are actually scored, LLM-as-judge, arena ratings, and pass@k.

Part 11 of 15

Inference, serving & efficiency

Sampling, the KV cache, prefill vs. decode, batching, and quantization.

Part 12 of 15

Guardrails, safety & deployment

Refusal training, layered defenses, and staged release.

Part 13 of 15

Long-context & context extension

Stretching RoPE, long-context data, best-fit packing, and why extension is its own late stage.

Part 14 of 15

Tool use, function calling & agents

OpenAPI tool specs, environment roles, real vs simulated trajectories, and how tool use is evaluated.

Part 15 of 15

Post-training at scale

The three-stage recipe, delta learning, upgraded GRPO, domain verifiers, and RL systems.

Reference

Glossary & numbers to know

Every bolded term across the series, linked back to where it's introduced, plus a quick-reference numbers card.

Already know some of this? Start here

If you...	Start at
Have never touched an ML paper	Part 1 — What is a language model
Know tokenization/embeddings, want the architecture	Part 3 — The transformer, block by block
Know transformers, want the data/training-systems side	Part 4 — Building the pretraining dataset
Have a base model, want to fine-tune it yourself	Part 7 — SFT & PEFT (LoRA)
Know SFT/RLHF, want reasoning-model RL specifically	Part 9 — Reasoning & RLVR
Just want to know how to read a benchmark number	Part 10 — Evaluation
Care about running models cheaply, not training them	Part 11 — Inference, serving & efficiency
Finished Part 11 and want serving as an engineering discipline	The sequel — LLM Serving, Interactively
Want the calculus behind backprop and gradient descent taught from scratch	Calculus in Motion, Interactively
Have a base model but it only knows a few thousand tokens	Part 13 — Long-context & context extension
Want to build tool-calling agents or MCP integrations	Part 14 — Tool use, function calling & agents
Want the full frontier post-training recipe end to end	Part 15 — Post-training at scale

Start from the beginning, or jump straight into whichever part you need. Start with Part 1 →