Jonathan Padilla

> Here's a first draft of a fine-tuned model I trained: [pogzyb/markuplm-phish](https://huggingface.co/pogzyb/markuplm-phish). From my experience, a fine-tuned MarkupLM performed better than a fine-tuned BERT on phish/malicious website classification. The final...

> Thanks for the PR! Would you mind separating out the example app from the core quantization logic? We can possibly integrate the example app at a later stage, but...

> This is awesome! Maybe a little more explanation in the docs would be good. Like what's the (performance) difference between binary and ubinary [here](https://huggingface.co/docs/transformers.js/main/en/api/utils/tensor#utilstensorquantizeembeddingstensor-precision--code-tensor-code) and what's the default right...

> Thanks for your reply. Found the related line [here](https://github.com/xenova/transformers.js/pull/683/files#diff-161ff30745072fdddb33e89b10d0c6ca2bb115bd389dd7499ad11a486ebec2abR1210): > > ```js > if (!['binary', 'ubinary'].includes(precision)) { > throw new Error("The precision must be either 'binary' or 'ubinary'"); >...

If we get this, my app for local rag would be solved. Although am concerned about data persistence, my use case is users upload textbooks and can perform vector similarity...