Model Genome: Fingerprint analysis to determine if LLMs are trained from scratch or derived

Model Genome is a reproducible analysis pipeline that verifies whether an LLM was trained from scratch or derived from an existing open-weight model, using only public data. The analysis covers three axes: architecture, tokenizer, and weights, combining the results into a single model 'genome'.

The first axis is the architecture fingerprint from config.json. It compares items such as model_type, vocab_size, hidden_size, intermediate_size, number of layers, and number of attention heads, specifically checking if the tuple (hidden_size, intermediate_size, layers, heads, kv heads) exactly matches external open-weight models. A complete structural match suggests a high likelihood that the architecture was adopted.

The second axis is tokenizer vocabulary overlap, and the third is CKA analysis of embeddings. However, row-wise embedding cosine similarity was invalid due to rotation invariance, and CKA also failed to clearly distinguish between continued pretraining and from-scratch training. Therefore, configuration files and tokenizers remain the core evidence.

The researchers applied the same criteria to public foundation models from nine Korean institutions, and the analysis results can be viewed in the Model Genome Korea demo.