Pixel LLM
A game about building a language model: ten stages from data collection to launch, where upstream decisions carry through to a final model card.
The pipeline is the game
Most people meet a language model as a chat box. Everything that decides how it behaves happened earlier: which data was collected, what was filtered out, how text was cut into tokens, how large the network is, what it was rewarded for.
Pixel LLM makes those ten stages playable and, more importantly, makes them depend on each other. A corpus cleaned carelessly in stage two produces noisier patterns in pre-training and a blurred view of the questions at evaluation. A small corpus cannot feed a large architecture. The player learns the pipeline by living with its consequences.
Ten stages, one model card
Stages do not know about each other. Each receives the model's profile, returns a validated result, and the run controller applies it. That contract is what lets a decision in stage one still matter in stage ten.
How a run is shaped
Data, cleaning, tokeniser, architecture, pre-training, evaluation, SFT, RLHF, red team, launch.
Intern is correct but simple, Researcher adds real tensions, Frontier stays close to real practice.
Short enough to replay, so a different decision can be tried the same evening.
Generated from the stats and traits the run accumulated; each stage closes on one of at least three archetypes.
Game logic that runs without the engine
Rules, formulas and the economy live in engine-independent code with no scene dependencies, so they can be unit-tested and rebalanced without launching the game. Scenes handle presentation and input only.
- Balance values sit in JSON tables validated against a schema at load, not in code.
- Every stage extends one base contract and returns a result that validates itself at its boundary.
- Saves are written atomically with an automatic backup and a migration hook.
- All player text goes through translation keys from the first day, English first.
Trailer and a minute of play
Captured from the current internal build.
In development under a working title. Internal web and Windows builds exist, external playtesting comes next, and no store page is live. The repository is private, so nothing here links out. The game teaches the pipeline; it does not train a real model.