Long document Q&A is not just about feeding all the text into the model. The official introduction of LensVLM-9B proposes to first view the compressed page image, and then select the page that needs to be expanded according to the question. This is a research project on the 9B visual language model. For entrepreneurs who are concerned about data retrieval, it provides another way to observe "find evidence first, read details later", but permission must be checked before application.
The page image must correspond to the original content
The official code sample contains tools to read the content of a specified page and provides a process for generating page images from text. When understanding this type of approach, you should pay attention to how page numbers are organized into the data: the same number may refer to different content after rearrangement, merging documents, or adding appendices.
In research comparisons, you can fix a set of allowed data, save the original text, rendered version and page correspondence table, and then record which pages were selected by the model. If the answer content is correct but the page number points to irrelevant material, it is still not suitable for making a traceability report based on it. Input preparation is also part of the evaluation, and you cannot just compare the last sentence.
Select evidence and answer questions to check separately
The code repository provides paths for selecting pages and answering assessments. ChinaOPC recommends separately recording “whether the necessary pages were found” and “whether the page content was used correctly”. For example, the answer needs to be supported by both the main text and the appendix. If you only read one of them, you may miss the qualification. Even if the text is expressed fluently, it should be marked as incomplete evidence.
You can also add samples where the data does not have answers, and observe whether the process still forces a certain page to be selected to give a positive conclusion. Research records also retain the number of expansions, source of results, and manual judgment; stronger compression or fewer reads cannot directly equate to better service quality.
Model weights and code permissions should be viewed separately
The model uses the Apple Machine Learning Research Model License, and the code uses a separate Apple Sample Code License. The original text of the model license states: “Research Purposes” does not include any commercial exploitation, product development or use in any commercial product or service.
This means that reading research materials cannot be confused with using that weight for product development and paid services. Please check the complete license through the source portal. Do not understand that the model is downloadable, the code is available, or a third party provides a demonstration as commercial use is permitted. This article does not provide a commercial deployment solution, nor does it reproduce its evaluation. Further reading: Why code and model weight licenses need to be checked separately.