yuqiannemo/llm-dna

LLM-DNA extracts low-dimensional behavioral signatures from language models by probing them with generated or dataset-backed text, then encoding model responses or hidden-state embeddings into DNA vectors. The primary workflow is exposed through the Python API and CLI, with support for single or parallel model runs, Hugging Face/local and hosted providers, metadata, response caching, and saved signatures. The README also describes comparing models and building evolutionary trees; the sampled files do not establish the implementation or wiring of tree construction.