Post training a model to explore Demoscenes
Summary
The author describes building and training Demoscener, a model designed to autocomplete and remix tiny JavaScript demos called Dweets. They collect about 25,000 Dweets from Dwitter.net, continue pretraining Qwen2.5-Code-7B, and reduce perplexity substantially. They then improve the model further with a curriculum based on Dwitter awesomeness scores and with RLHF-style preference training using their own aesthetic judgments. The final system lets the author generate multiple continuations from the same prefix, compare them visually, and branch into new creative directions. The post ends by sharing Demoscener-7B and the dataset on HuggingFace.
AskAI Classifications
Sectors
No sectors detected
Functions
Developer and IT Infrastructure
Development Platforms
Linked Companies
Hugging Face, Inc.
$10M to $25M