Official Codebase of 3DCodeVerse
Python
eval/ — evaluating a bare LLM / VLM on 3D coding
llm/ asks a model one-shot — no plugin, no tools in the loop — over 3DCodeBench, the held-out
dialect sets and ten prompt batteries (llm/batteries/*.yaml); every answer is executed and scored with
geometry and image metrics against references. docs/ holds the earlier evaluation write-ups
(EVAL.md, COMPLEXITY.md, PAPER_WRITING.md); the harness benchmark they describe (eval/bench) ran
on the v0.0.x CLI harness and lives in those releases.
cd eval
python -m llm.config # what the evaluation can / cannot find
python -m pytest llm/tests # offline tests
Run data is never in git: llm/data/prompts/*.jsonl is built by python -m llm.build_prompts from the
gated Hub assets.