# LLM Training Mastery

An intensive, project-based course for learning to build, train, evaluate, and improve language models at home, with an AI tutor and human review.

**Target:** gain hands-on deep-learning training expertise, centered on language models, with a possible later path into LLM research engineering. The course is designed to produce inspectable evidence of ability: working implementations, controlled experiments, reproducible reports, and contributions other researchers can use.

**Start with [Choose your learning route](docs/15-learning-paths.md).** Developers new to ML can take the prerequisite bridge; those with basic ML can work through the core modules; experienced engineers can use the optional [24-week intensive plan](docs/00-personalized-bootcamp.md). Advance through evidence of understanding at a sustainable pace. Timeline estimates are planning tools, not mastery or hiring guarantees.

## Open the interactive workbench

Open the [hosted course](https://llmmastery.org/), or clone the public repository and start the local workbench:

```bash
git clone https://github.com/victusfate/llm_mastery.git
cd llm_mastery
npm ci
npm run dev
```

Visit **http://127.0.0.1:8765/site/** for 35 narrated submodules, 52 dedicated lab guides, formatted readings, concept popups, interactive numerical experiments, randomized theory checks, scheduled review, and hands-on tutor handoffs. No API key or cloud spend required. [Workbench guide](docs/10-interactive.md).

## Start here

1. Read the [starting instructions](START_HERE.md).
2. Complete the [placement diagnostic](assessments/00-placement.md) before choosing a route.
3. Follow the [24-week bootcamp](docs/00-personalized-bootcamp.md), or the [fuller curriculum](docs/02-curriculum.md), using the linked module guides.
4. Use the [tutoring protocol](docs/03-working-together.md) at each session.
5. Track evidence using the [mastery rubric](assessments/01-mastery.md) and [progress tracker](progress/README.md).

## What you will build

| Artifact | Demonstrates |
| --- | --- |
| A language model training stack built from basic tensor operations | Autograd, tokenization, decoder architecture, numerical correctness |
| A reproducible pretraining study | Data quality, optimization, checkpointing, scaling, evaluation |
| A measured systems improvement | Profiling, kernels, distributed training, performance tradeoffs |
| A post-training comparison | SFT, preference learning, PPO, verifiable-reward RL, robust evaluation |
| An original specialization study and an external contribution | Research judgment, ownership, collaboration, technical communication |

The initial release includes a working learning interface, course instructions, and assignment specifications. Learners implement training code during the course; no training experiment results or learner achievements are claimed yet. This is an independent curriculum, not an accredited program or an affiliation with any linked institution.

## Course map

| Guide | Purpose |
| --- | --- |
| [24-week bootcamp](docs/00-personalized-bootcamp.md) | Primary experienced-engineer route |
| [Interactive workbench](docs/10-interactive.md) | Lessons, visuals, quizzes, learning records |
| [Home hardware](docs/09-home-lab.md) | Personal hardware settings, CPU/GPU routes, and measured memory limits |
| [Open-source lane](docs/08-open-source.md) | Current contribution candidates and first-ten-day plan |
| [Timeline](docs/01-timeline.md) | Prerequisites, minimum plausible routes, hours, reforecasting |
| [Curriculum](docs/02-curriculum.md) | Every week, deliverables, module dependencies |
| [Working together](docs/03-working-together.md) | Tutor prompts, independent work, review cadence |
| [Compute and setup](docs/04-compute.md) | CPU fallback, GPU tiers, cost accounting, environment setup |
| [Portfolio projects](projects/README.md) | Experiments and acceptance criteria |
| [Hiring strategy](docs/05-career.md) | Job evidence, interviews, visibility, exceptional-offer limits |
| [Reading library](docs/06-resources.md) | Primary papers and official implementation references |
| [Experiment standards](docs/07-experiments.md) | Fair comparisons, leakage controls, reproducibility |
| [Prerequisite bridge](modules/00-bridge.md) | Extra preparation when the diagnostic reveals gaps |
| [Contributing](CONTRIBUTING.md) | How others can use and improve this public course |

**Schedule:** a 12-hour daily window on five days, containing 9 hours of coursework and 3 hours for meals, exercise, and breaks; 7 coursework hours Saturday; Sunday off. That is 52 coursework hours per week, not 84 effective learning hours. Adapt to sustained performance rather than chasing logged hours.

**Compute:** customize the [home-lab plan](docs/09-home-lab.md) for your browser, CPU, Apple Silicon, or GPU hardware and optional cloud budget. Start with $0 required cloud spend; scale experiments to measured memory and runtime. The [two-PC example](docs/09-home-lab-example.md) is an optional case study.

**Progress:** learning records, lab notes, submodule notes, and hardware preferences stay in your browser. Export/import JSON to move them between devices; bookmarks save your page location. See [Saving progress](docs/17-saving-progress.md).

**Contribute immediately:** the [open-source lane](docs/08-open-source.md) targets a small reproduction in Week 1 and a reviewable patch within two weeks, beginning with LM Evaluation Harness or DataTrove. Acceptance remains up to maintainers.

**Sources checked:** September 8, 2026. The roadmap combines original assignments with selected primary references. Hiring requirements and software APIs change; refresh them at the scheduled checkpoints. Original repository material uses the [MIT license](LICENSE.md); external materials retain their own terms.

## Develop the course

Interactive source is TypeScript in `src/site/`. Browser modules are generated by `npm run build`; edit the TypeScript source rather than generated `.mjs` files. Python remains the language for numerical examples and training exercises. Run `npm run check` for type checks, course tests, Markdown links, and the static build. `npm run test:browser` uses a separately installed Playwright and a running local preview.

The repository uses [Scaffold](https://github.com/victusfate/scaffold)'s shared development harness. Update it with `bash bin/sync-from-scaffold.sh`; `.scaffold-keep` protects course-owned configuration. [Platform decisions](docs/course-platform/design.md), [learning routes](docs/15-learning-paths.md), [tutor interfaces](docs/16-tutor-interfaces.md), and [hosting](docs/14-hosting.md) describe the current design.

## Deployments

Pushes to `main` run the [GitHub Actions workflow](.github/workflows/pages.yml): validation and a static build, followed by deployment jobs for the GitHub Pages mirror and Cloudflare Pages (`llmmastery.org`). Pull requests run validation only. Cloudflare deployment requires the repository secrets `CLOUDFLARE_API_TOKEN` (Account → Cloudflare Pages → Edit, limited to the hosting account) and `CLOUDFLARE_ACCOUNT_ID`. Both secrets are configured for this repository. Forks must configure their own hosting credentials; missing credentials produce an explicit setup error.

For an authenticated local production deployment, run `npm run deploy:cloudflare`. See [hosting and credential setup](docs/14-hosting.md).

## License

Original course code and documentation are available under the [MIT license](LICENSE.md). You may use, adapt, and redistribute them, including commercially, while retaining the license notice. Linked papers, videos, datasets, models, and other third-party materials retain their own licenses.