Skip to content
PostTrainLLM docs
Esc
↑↓navigate↵open⌘Jpreview
On this page

PostTrainLLM learning curriculum — ground up

This is the ground-up roadmap from first principles to a self-improving factory for Mac-local specialist models. The owner’s active route for 28 September–27 December 2026 is the 13-week inference-systems sprint; use this curriculum when that route calls for a foundation or remediation lesson.

The goal is not to “finish a course.” The goal is to build durable taste:

  • know what a model is actually doing,
  • know what changes when data, loss, optimizer, architecture, or eval changes,
  • know when a post-training result is real,
  • and know how to turn a failure into the next better recipe.

10/10 Bar

The roadmap is good enough only if it satisfies all of these:

  1. Ground-up order — no DPO, LoRA, RLVR, or eval jargon before the underlying model/training concepts are in place.
  2. One spine — every module fits the same arc: concept -> toy implementation -> posttrainllm anchor -> factory consequence.
  3. Exercises, not passive reading — each module has something to compute, inspect, run, or write.
  4. Mastery gates — each module has a concrete “you understand it when” check.
  5. Project transfer — every module says how it changes SQL/factory work.
  6. External anchors — one or two canonical external resources per stage, used as support, not as the curriculum owner.
  7. Progress tracking — actual interactive state lives in the browser-local workspace, with evidence; the manual record is a historical reference.
  8. Cadence — each week ends in a note, a repo artifact, or a recipe change.
  9. Failure feedback — failed posttrainllm runs create the next learning prompt.
  10. No random walk — interesting topics are parked unless they improve target selection, data, post-training, eval, packaging, or reporting.

Operating Loop

Start at the single Next Session card. Choose one module and one exercise; record one explanation in the owner’s own words, actual work, errors, and the connection to a repo artifact.

read -> exercise -> explain -> immediate check -> delayed recall -> verified

The Learning Loop owns status rules and recall scheduling. After the immediate gate passes, the next module may start while recall checks remain due at +2 and +7 days. Complete those checks closed-book using changed examples. Repair failed recall before advancing further. Reading a file or generating a lesson is not evidence of mastery.

A typical 30-minute session gives 5 minutes to due recall, 10 to focused reading, 10 to the exercise, and 5 to the checkpoint. Split substantial labs across sessions. All existing courses, sessions, paths, articles, and artifacts remain reference material; adding a source does not change the current module.

Canonical External Anchors

Use these as supporting material:

The project still owns the path. External material fills intuition gaps; it does not decide what to build next.

Master Roadmap

# Module Learn Exercise Repo Anchor Mastery Gate
1 Functions, data, parameters A model is a parameterized function; learning means choosing parameters from data Fit y = mx + b by hand on 5 points; explain fixed data vs moving parameters session-01-neural-net-basics.md You can explain what a “parameter” is without using LLM examples
2 Loss and gradient descent Loss turns wrongness into one number; gradients say how to change parameters Compute MSE for two (m,b) guesses; take one gradient-descent step session-02-gradient-descent.md You can predict what too-high and too-low learning rate look like
3 Vectors, matrices, tensors Neural nets are mostly structured multiply/add over arrays Rewrite a one-input line as a dot product; trace tensor shapes through one layer session-09-tensors.md, essential-vs-optimization.md You can read a shape error and identify which axis is wrong
4 Non-linear neural nets + backprop Stacking linear layers only stays linear; activations and chain rule make depth useful Train a tiny 2-layer MLP on a non-linear toy dataset; explain backprop as credit assignment session-03-non-linearities.md You can explain why a model can fit curves after adding activation functions
5 ML paradigms and scaling Supervised learning, self-supervision, imitation, RL, and scale each solve different parts Classify posttrainllm attempts as pretrain, SFT, preference tuning, eval, or routing session-04-ml-paradigms.md, session-05-scaling.md You can say why scale helps knowledge but does not fix bad evals or bad data
6 Tokenization, embeddings, language modeling Text becomes tokens; tokens become vectors; next-token prediction creates language skill Tokenize three prompts; inspect how SQL punctuation and identifiers split session-06-tokenization-embeddings.md, ../tool_call_extractor.md You can explain why tokenization affects SQL/tool-call reliability
7 Attention and transformer blocks Attention routes information across positions; transformer blocks repeat attention + MLP Work one tiny attention example with query/key/value vectors and shapes session-10-attention.md, llm-mechanics-fundamentals.md You can describe what attention can copy/route that an MLP alone cannot
8 Training mechanics Batches, epochs, optimizers, schedules, precision, overfit checks, and loss curves govern whether training worked Overfit a tiny dataset or inspect an existing overfit gate; identify failure mode from a loss curve session-08-training-mechanics.md, ../training_guide.md You can tell data bug vs LR bug vs capacity bug from symptoms
9 Post-training: SFT, LoRA, preference tuning SFT teaches behavior; LoRA changes a low-rank slice; DPO/SimPO shape preferences and can collapse Compare successful SQL SFT vs failed hygiene SimPO; run/inspect LoRA geometry ../training/sft.md, ../training/dpo.md, ../factory/lora-geometry.md You can explain why the hygiene SimPO run collapsed without hand-waving
10 Evals, rewards, and self-improvement Frozen evals, verifiable rewards, traces, failure taxonomy, and public reports make improvement measurable Build/inspect SQL candidate-selection rows; attach slice metrics and trace review to a run session-11-evals-rewards.md, ../factory/eval-protocol.md, ../techniques/sql-technique-backlog.md You can design the next SQL recipe with target, data, reward/eval, stop rule, and report fields

Where Existing Sessions Fit

All ten modules now have a polished session. The original eight are the foundation half; the three sessions added later (9, 10, 11) fill the gaps that were previously reference-only.

Session Roadmap Module
session-01-neural-net-basics.md 1
session-02-gradient-descent.md 2
session-09-tensors.md 3
session-03-non-linearities.md 4
session-04-ml-paradigms.md 5
session-05-scaling.md 5
session-06-tokenization-embeddings.md 6
session-10-attention.md 7
session-08-training-mechanics.md 8
session-07-behavior-learning.md 9
session-11-evals-rewards.md 10

Session file numbers are creation order, not reading order — read by module. The three later sessions (9 = tensors → Module 3, 10 = attention → Module 7, 11 = evals/rewards → Module 10) closed the previously-missing gaps.

Coverage Beyond the Spine

This curriculum is the ground-up spine. For the guarantee that every shipped subsystem in the project — post-training internals, quantization, serving, interpretability, WASM/WebGPU, VLM, the factory loop — has a learning anchor, see coverage-map.md. The spine teaches you to read the system; the coverage map is the index that proves nothing was left unread.

Applied Industry Case Studies

Use the industry learning roadmap after the relevant foundations. These are learning exercises, not new required modules:

  • Savante / Aryabhata, after Modules 9–10: reconstruct a specialist recipe, audit data and reward design, and distinguish reported gains from causal proof.

  • Bonsai 2 27B, in the quantization and runtime paths: distinguish compressed weights from runtime cost and design a fair capability-per-RAM comparison.

  • QORL, after Module 10 in evaluation and runtime study: compare offline plan search and parameter routing with native PostgreSQL planning, including held-out measurements, regressions, and tuning-cost break-even. The retained owner PRD excludes LLM training.

  • Inside vLLM, after transformer foundations in the inference/runtime path: trace scheduling and KV-block ownership, predict latency/throughput tradeoffs, and map the single-Mac-to-cluster boundary.

  • Splash, after transformer and training-mechanics foundations: study how per-model Metal kernels, speculative drafts, cache policy, and memory planning trade generality for Mac-local serving speed, then design a fair same-Mac comparison.

Each case study has source links, an exercise, and a mastery gate. Track the checkpoint in learning progress. No model run is needed to complete the initial source audit.

Current Starting Point

For the active inference sprint, start with its Day 1 diagnostic. If a foundation gap appears, use the smallest matching module below and retry with a changed example. For a complete ground-up pass, start at Module 1 unless the owner can pass its mastery gate out loud.

SQL candidate selection is retained lab material, not an active experiment. The learning path does not jump straight there. The correct bridge is:

parameters -> loss -> gradients -> tensors -> neural nets
-> tokens -> transformers -> training loops
-> SFT/LoRA/DPO -> evals/rewards/self-improvement

The SQL factory is the lab. Ground-up understanding is the curriculum.

Checkpoint Template

Use the canonical checkpoint template after each exercise. Record the owner’s explanation, actual work, repo connection, immediate result, dated recall checks, and unresolved confusion. Keep old checkpoints so progress and repairs remain auditable. Experiment reports and learning checkpoints serve different purposes; a new run is not required to verify a foundational concept.

Was this page helpful?