Old maps (and perhaps new ones) used to add fake little alleys so a publisher could quickly spot publishers infringing on their IP rather than going out and actually mapping. I wonder if something similar is possible with LLMs.
replies(6):
实际上，对于后续训了很久很久的这个模型，Honestagi能够分析出这个量级的相似性我已经很诧异了，因为这个模型为了续训洗参数，所付出的算力甚至早就足够从头训一个同档位的模型了。听同事说他们为了洗掉千问的水印，采取了不少办法，甚至包括故意训了脏数据。这也为学术界研究模型血缘提供了一个前所未有的特殊模范吧。以后新的血缘方法提出可以拿出来溜溜。
In fact, I'm surprised that HonestAGI's analysis could show this level of similarity for this model that had been post-trained for a long time, because the computing power used to train-wash the parameters of this model was enough to train a model of the same size from scratch. I heard from my colleagues that they took many measures to wash off Qwen's watermark, even deliberately training on dirty data. This also provides an unprecedented case study for the academic community studying model lineage. If a new lineage method is put forward in the future, you can take it for a spin.