A Hugging Face community developer built and published a pipeline that checks whether a large language model was actually trained from the ground up or assembled on top of someone else’s open-weight release. The project, posted under the handle Proto_AGI and titled Model Genome, examines three parts of any public model: its architecture settings, its tokenizer, and its weights, then reduces the comparison to a single readable label. The author applied the identical method to nine South Korean AI companies’ public foundation models and posted the full results on a companion Hugging Face Space.
Claiming a model was trained from scratch has become a marketing line with money behind it. National AI initiatives fund domestic labs partly on the premise that they are building independent foundation models rather than repackaging Qwen or Llama, and venture rounds get priced on that same premise. Until now, checking that claim meant taking a company’s word for it, because no outsider had assembled a repeatable, public method for testing it.
The first check reads a handful of numeric settings, such as hidden layer width, layer count, and attention head counts, straight out of the configuration file every model ships with. Those numbers combine into a shape. When a new model’s shape lines up exactly with an existing open release, such as Qwen2.5-72B or Llama-3.1-8B, coincidence stops being a credible explanation once enough fields agree simultaneously.
The second check compares vocabulary files rather than settings, measuring how much of one model’s token set overlaps with another’s. Near-total overlap signals a tokenizer copied wholesale. Partial overlap paired with a matching architecture is what the author calls a foreign frame carrying a locally built vocabulary: a team can adopt someone else’s blueprint while still training its own language layer from scratch.
The hardest check compares the weights themselves, and it is where the project earns its credibility, because the author documents two dead ends before landing on something that works. Comparing embedding vectors token by token failed outright, since a neural network’s internal representation space can be rotated without changing what it encodes, so two closely related models can look unrelated by that naive measure. A more advanced similarity metric, Centered Kernel Alignment, fixed that false negative for genuinely independent models, returning a near-zero score against unrelated bases. It could not reliably flag derivative models, though: a model that had been continued-trained on a foreign base landed only slightly above what two unrelated but similarly built models score against each other.
That gap is why the author treats configuration and tokenizer evidence as the primary signal and the weights check as supporting evidence rather than a verdict on its own. Applied to the nine companies, the results were not uniform. Some lined up precisely with an outside base on both architecture and tokenizer. Others showed no foreign match at all. Several sat in between. The author is explicit that a match is not proof of wrongdoing: building on an open-weight base from Qwen, Llama, DeepSeek, or Mistral is common practice across the industry, and the tool is designed to surface lineage, not to accuse anyone of copying.
What changes is who gets to run the check. Before this pipeline, confirming a from-scratch claim meant trusting a lab’s own disclosure or waiting for a competitor’s leaked teardown. Now the underlying method works against any two public repositories on the Hub, which means a journalist, a rival lab, or a fund writing a term sheet can test a from-scratch claim before money moves on the strength of it.
Published on Hugging Face by its author, Proto_AGI, as a community blog post rather than an official Hugging Face announcement.
