modelDNA：基于采样权重指纹的校准谱系验证与合并分解

Jul 16, 2026

Open-weight language model lineage is often self-reported: Hugging Face’s base_model metadata is optional and frequently absent, making provenance disputes hard to resolve and forcing manual re-analysis. modelDNA fingerprints a model using only ~100–300MB of HTTP byte-range reads, compares it to a reference database across four signal families, and outputs one of eight calibrated verdicts with abstention to avoid false accusations. On 15 Hub models with org-documented ancestry, it reaches AUROC 1.0 and correct top-1 attribution. It also derives merge decomposition by recovering mixture weights from fingerprints via constrained least squares. Fingerprints, benchmarks, and an inferred lineage graph for 55 models are publicly reproducible offline.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages13

Reading Time33 min

Introduction

Lineage metadata limitations

Why lineage disputes matter

Existing methods and the missing tool

Design asymmetry: avoiding false accusations

Contributions

Fingerprint design based on published signal families

Calibrated verdict engine with abstention

Merge decomposition from fingerprints

modelDNA: Calibrated Lineage Verification and Merge Decomposition from Sampled Weight Fingerprints Muhammad Awais Bin Adil · SaadAamir Independent · [binadilawais@gmail.com](binadilawais@gmail.com) · [saadaamir473@gmail.com](saadaamir473@gmail.com)[ ](saadaamir473@gmail.com)[v1.0](v1.0), July 2026 Code: [https://github.com/AwaisAdilKhokhar/modelDNA](https://github.com/AwaisAdilKhokhar/modelDNA) · Live scanner: [https://huggingface.co/spaces/AwaisAdilKhokhar/modelDNA](https://huggingface.co/spaces/AwaisAdilKhokhar/modelDNA) · Data: [https://huggingface.co/datasets/AwaisAdilKhokhar/modeldna-atlas](https://huggingface.co/datasets/AwaisAdilKhokhar/modeldna-atlas) Abstract The lineage graph of open-weight language models is self-reported: Hugging Face’s base_model metadata field is optional and unverified, and over 60% of Hub models document no parentage at all. Methods for detecting lineage from weights exist in the research literature, but each ships as paper code tied to one signal and one experiment; when a provenance dispute breaks, the analysis is redone by hand. This report describes modelDNA, a tool that fingerprints a model from roughly 100-300 MB of ranged [HTTP reads](HTTP reads) (instead of a full 15 GB download for a 7B model), compares the fingerprint against a reference database of foundation models across four published signal families, and returns one of eight verdict classes with a calibrated probability, preferring honest abstention to confident error. On a benchmark of 15 real Hub models with org-documented parentage, judged against 8 candidate bases (13 positives, 107 hard negatives), the system achieves AUROC 1.0, zero false positives at its reporting threshold, and 13/13 correct top-1 parent attribution. The report’s second contribution is merge decomposition. Every mainstream weight-merging method is (near-)linear per tensor, and fingerprint sample positions are deterministic functions of tensor identity, so a merged model’s fingerprint is the same linear combination of its parents’ fingerprints. Mixture weights can therefore be recovered from fingerprints alone by sum-to-one constrained least squares, which is algebraically a well-conditioned regression in task-vector space. Against merges with published mergekit configurations as ground truth, the method recovers a slerp merge’s layer-interpolation curves at r = 0.999 and a dare_ties merge’s mixture weights to within 0.011 of the published values, without downloading any weights beyond the fingerprints. All fingerprints, benchmarks, and the inferred lineage graph of 55 models are public and reproducible offline. 1. Introduction Hugging Face hosts nearly three million model repositories (2 . 9M as of July 2026), and the graph of who fine-tuned whom is almost entirely on the honor system. The base_model metadata field is optional, unverified, and frequently absent: Horwitz et al. [5], building their Model Atlas, measured that more than 60% of Hub models carry no documented parentage, and had to manually inspect the largest connected components of their own dataset just to fill in missing values. Hugging Face itself ran a volunteer sprint asking the community to open metadata PRs by hand. This matters beyond bookkeeping. Three well-documented incidents shaped this project. In July 2025, an anonymous group published an analysis showing a 0.927 correlation between the per-layer attention-projection standard-deviation curves of Huawei’s Pangu Pro MoE and Alibaba’s Qwen2 . 5-14B, against an unrelated-pair background of roughly 0.3-0.7 [1]; Huawei denied derivation, the analysis repo was taken down, and the dispute ran in the international press with no neutral party able to rerun the numbers. In September 2024, Reflection-70B was marketed as a breakthrough fine-tune until community weight-diffing showed the claims didn’t hold up. In June 2024, the Llama3-V project was shown to be substantially copied from MiniCPM-Llama3-V 2.5. In each case the i

FAQs

modelDNA如何在不完整下载权重的情况下对模型进行指纹识别？

modelDNA通过约100–300MB的HTTP字节范围读取生成指纹，而不是下载完整的7B权重。指纹采样位置由确定性函数生成，从而保证不同模型之间可直接比较。

当证据不足时，modelDNA如何避免错误指控？

系统采用保守阈值与“弃权区间”，并在输出时提供校准概率。背景分布用于对缺失证据进行不偏倚处理，优先选择诚实弃权以降低误报风险。

modelDNA的合并分解（merge decomposition）方法依赖什么关键性质？

主流权重合并对每个张量近似线性；同时指纹采样位置是张量身份的确定函数。因此，合并模型的指纹等于父模型指纹的同一线性组合，可用满足和为1约束的最小二乘在指纹空间回归合并混合权重。

Jul 16, 2026

Upload by Evangeline