A Hugging Face community article published August 8, 2026 by the user Proto_AGI (mayafree) describes a reproducible method for testing whether a released LLM was pretrained from scratch or derived from an existing open-weight base. The method fingerprints three axes: architecture fields in config.json (hidden_size, intermediate_size, num_hidden_layers, attention and KV head counts), tokenizer vocabulary overlap from tokenizer.json using min-set overlap, and embedding-space linear CKA. The authors report that row-wise embedding cosine is uninformative because of rotational invariance, and that CKA reliably confirms from-scratch training (near-zero) but poorly detects derivation (about 0.25 for a continued-pretrained model versus about 0.21 between unrelated models), so config plus tokenizer remain primary evidence.

• huggingface.co2026-08-08