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

Learning Pipeline

Learning Pipeline

The owner’s learning track is now explicitly ground-up. It still reinforces the posttrainllm factory, but it does not start with post-training jargon.

Primary path:

math intuition -> tiny neural net -> training loop -> transformer
-> LLM behavior -> post-training -> evals -> self-improving factory

Learning is not separate from building; every module should eventually make the next run better.

Progress is tracked in learning-progress.md.

Principle

concept -> toy implementation -> posttrainllm anchor -> recipe improvement

If a learning topic does not improve target selection, data, post-training, eval, packaging, or reporting, park it until it does.

Ground-Up Master Roadmap

The canonical roadmap is docs/learn/curriculum.md.

It defines the 10-module path:

  1. functions, data, parameters
  2. loss and gradient descent
  3. vectors, matrices, tensors
  4. non-linear neural nets and backprop
  5. ML paradigms and scaling
  6. tokenization, embeddings, language modeling
  7. attention and transformer blocks
  8. training mechanics
  9. post-training: SFT, LoRA, preference tuning
  10. evals, rewards, and self-improvement

The SQL factory is the lab, not the starting point. Do not skip the foundation unless the module mastery gate can be passed out loud.

All ten modules now have a polished session. For the guarantee that every shipped subsystem (post-training internals, quantization, serving, interpretability, WASM/WebGPU, VLM, the factory loop) also has a learning anchor, see docs/learn/coverage-map.md.

Factory-Attached Learning Sequence

Order Topic Why Now Project Work It Unlocks Primary Docs
1 Eval design Bad evals create fake progress frozen gates, slice metrics, leakage checks, regression gates docs/factory/eval-protocol.md, docs/learn/eval-methodology-2026-06-08.md
2 Data for post-training Most improvement is data/reward shape, not optimizer novelty SFT rows, preference pairs, candidate sets, synthetic data, trace-to-data docs/factory/post-training-factory.md, docs/recipes/from-traces.md
3 SFT + LoRA mechanics Current SQL wins came from SFT/LoRA; failures need adapter diagnosis rank/layer sweeps, LoRA geometry, overfit checks docs/training/sft.md, docs/lora_guide.md, docs/factory/lora-geometry.md
4 Preference tuning Hygiene SimPO collapsed; need to understand why reference-anchored DPO retry, length-balanced negatives, KL/reference anchoring docs/training/dpo.md, docs/techniques/method-vs-recipe.md
5 Verifiable rewards SQL/tool-calling improvement needs executable rewards SQL execution reward, BFCL AST reward, unit-test rewards docs/techniques/sql-technique-backlog.md, docs/learn/tool-calling-frontier-parity.md
6 RLVR / ReST / OAPL Only useful after reward surfaces are clean batch rollouts, offline scoring, policy lag, candidate selection docs/factory/batch-posttraining.md, docs/learn/advanced-llm-training.md
7 Failure analysis Failed runs must become data trace review, failure taxonomy, targeted retry data docs/factory/reports.md, docs/attempt-ledger.md
8 Public reporting Public artifacts are a product surface case-study reports, blockers, competition comparison, reproduction notes docs/factory/case-study-template.md, docs/factory/public-artifacts.md

Current Practical Curriculum

This is the project lab sequence. It should run alongside the ground-up path, not replace it.

Module 1 — SQL Eval Quality

Goal: know when a SQL improvement claim is real.

Do:

  • Read the SQL attempt ledger.
  • Compare exact match vs execution accuracy.
  • Study why public b-mc2 exact is not enough.
  • Build or run the Spider/BIRD-style execution gate when DBs are local.

Deliverable:

  • A short note in the next SQL report explaining which metric is trusted and which metric is only directional.

Module 2 — Candidate Selection

Goal: understand why “choose the best answer” can be easier than open generation.

Do:

  • Read docs/techniques/trainloop-teardown.md.
  • Inspect scripts/build_sql_candidate_choice.py.
  • Generate candidate-choice rows from existing SQL predictions.
  • Train/eval a small candidate-selection adapter only after the recipe card is written.

Deliverable:

  • Candidate-selection accuracy overall and by slice.

Module 3 — Preference Tuning Failure

Goal: explain the failed hygiene SimPO run without hand-waving.

Do:

  • Read the hygiene run report.
  • Compare ref-free SimPO vs reference-anchored DPO.
  • Inspect degenerate predictions and trace review.
  • Define a safer retry recipe.

Deliverable:

  • A recipe card that predicts how the retry avoids collapse.

Module 4 — Adapter Geometry

Goal: see whether adapter failures are capacity, targeting, or optimization problems.

Do:

  • Run scripts/lora_geometry.py on successful and failed SQL adapters.
  • Compare rank, stable rank, Frobenius norm, and top modules.
  • Attach lora-geometry.json to the next run.

Deliverable:

  • One paragraph in the report explaining whether the adapter update looked concentrated, diffuse, too small, or too large.

Learning Is Complete For A Topic When

  • the concept is explained in owner-readable language,
  • the explanation points to repo evidence,
  • it changes a recipe, eval, or report,
  • and the next run can verify whether that change mattered.

Was this page helpful?