Learning Progress Tracker

This tracker makes the owner learning pipeline measurable. The goal is to learn from the ground up while using posttrainllm as the lab.

Status values:

Ground-Up Roadmap Progress

Canonical roadmap: learn/curriculum.md. Coverage index (every subsystem → anchor): learn/coverage-map.md.

All ten modules now have a polished session; remaining work is mastery, not authoring. Status is per-module learning state, not doc-existence.

#	Module	Status	Evidence	Next Concrete Action
1	Functions, data, parameters	`reading`	Session 1 exists and has self-checks	Pass mastery gate out loud; write checkpoint
2	Loss and gradient descent	`reading`	Session 2 exists and has worked examples	Compute one MSE + gradient step by hand
3	Vectors, matrices, tensors	`reading`	Session 9 (tensors) written; anchors LoRA shape logic	Trace one layer’s shapes; read a shape error and name the wrong axis
4	Non-linear neural nets + backprop	`not-started`	Session 3 exists	Run/inspect tiny non-linear example
5	ML paradigms and scaling	`not-started`	Sessions 4 and 5 exist	Classify posttrainllm attempts by paradigm
6	Tokenization, embeddings, language modeling	`not-started`	Session 6 exists	Tokenize SQL prompts and inspect splits
7	Attention and transformer blocks	`reading`	Session 10 (attention) written; ties to interpretability heatmap	Work one tiny Q/K/V attention example
8	Training mechanics	`not-started`	Session 8 exists	Inspect tiny overfit gate and failure symptoms
9	Post-training: SFT, LoRA, preference tuning	`reading`	SFT/LoRA/DPO docs and SQL run evidence exist	Explain SQL SFT win vs SimPO collapse
10	Evals, rewards, self-improvement	`reading`	Session 11 (evals/rewards) written; eval protocol, attempt ledger, SQL candidate-choice tools exist	Build/inspect candidate-selection rows and report slice metrics

Module	Status	Evidence	Next Concrete Action
Eval design	`applied`	Frozen SQL gates, public-vs-synthetic distinction, slice metrics tooling	Add public Spider/BIRD execution gate when DBs are local
Data for post-training	`applied`	SQL SFT rows, preference pairs, failure-derived rows, candidate-choice builder	Build candidate-selection train/eval rows from existing predictions
SFT + LoRA mechanics	`applied`	Expanded synthetic SFT worked; public v4 worked on public exact; LoRA geometry tooling exists	Run controlled rank sweep only after next target is frozen
Preference tuning	`applied`	Hygiene SimPO collapsed and is documented	Write/run reference-anchored DPO retry recipe
Verifiable rewards	`reading`	SQL execution and BFCL AST matching are understood as target reward surfaces	Turn SQL candidate selection into a scored reward/data loop
RLVR / ReST / OAPL	`not-started`	Batch plan renderer exists; no model run	Start only after candidate-selection evidence exists
Failure analysis	`applied`	Failure taxonomy, trace review tooling, attempt ledger	Attach `trace_review.md` to every new SQL run
Public reporting	`applied`	Public artifacts registry, case-study template, publish-check	Re-render public SQL artifact with perf and public execution when available

The next project-lab focus is candidate selection for SQL:

The next ground-up focus is Module 1 -> Module 2:

A module reaches verified only when:

Reading alone is not enough.