godot-dodo

Why we included this project

Most finetuning recipes lean on a large model to write the training data, but godot-dodo does the opposite: it pulls human-written GDScript from MIT-licensed GitHub repositories, splits it into individual functions, and uses a language model only to label each snippet with a comment, producing comment:code pairs in the style of CodeSearchNet. The repo bundles the dataset assembly scripts, a pre-built 60k-row dataset, finetuning commands for LLaMA 7B and 13B, links to the resulting model weights, and a performance report. The report is the interesting part: the finetuned models are significantly more consistent than gpt-4 and gpt-3.5-turbo at generating valid GDScript syntax, though they tend to assume objects exist that the instruction never defined. Anyone building code generation for Godot, or anyone who wants a worked example of language-specific finetuning, can reuse the whole pipeline as-is.