[{"data":1,"prerenderedAt":134},["ShallowReactive",2],{"mainfooter":3,"guide/guides/assessing-data-transparency":42,"guides-landing":100},{"id":4,"title":5,"body":6,"description":34,"extension":35,"meta":36,"navigation":37,"path":38,"seo":39,"stem":40,"__hash__":41},"docs/footer.md","Footer",{"type":7,"value":8,"toc":30},"minimark",[9],[10,11,12,13,20,21,25,26,29],"p",{},"Supported by the Centre for Language Studies and the Dutch Research Council. Website design & development © 2024 by ",[14,15,19],"a",{"href":16,"rel":17},"https://www.bstn.nl",[18],"nofollow","BSTN",". This version of the index generated ",[22,23],"versionlink",{"repo":24},"data",", website content last updated ",[22,27],{"repo":28},"website",".",{"title":31,"searchDepth":32,"depth":32,"links":33},"",2,[],"Supported by the Centre for Language Studies and the Dutch Research Council. Website design & development © 2024 by BSTN. This version of the index generated , website content last updated .","md",{},true,"/footer",{"title":5,"description":34},"footer","JMbSG6ltlINXtOPTyikVPd3XOmp4GPFz_R5lyZZF_eA",{"id":43,"title":44,"author":45,"body":46,"date":92,"description":93,"extension":35,"image":94,"meta":95,"navigation":37,"path":96,"seo":97,"status":98,"stem":51,"__hash__":99},"guides/assessing-data-transparency.md","How to assess openness of training data","Dick Blankvoort",{"type":7,"value":47,"toc":90},[48,53,56,59,62,65,81,84,87],[49,50,52],"h1",{"id":51},"assessing-data-transparency","Assessing data transparency",[54,55],"author",{":author":54},[57,58],"date",{":date":57},[10,60,61],{},"Intuitively, \"data openness\" seems like a binary: a model either discloses its training data or it does not. In reality, several factors complicate this assumption.\nFirst, as reflected in our database, pretraining data and fine-tuning data are distinct. Pretraining relies on massive web scrapes to build latent conceptual representations, while fine-tuning shapes the model into a specific role (e.g., a helpful assistant). Often, a model's fine-tuning data is public while its base data remains proprietary. Because of this, we evaluate these sources separately.",[10,63,64],{},"Second, true openness requires transparency across the entire fine-tuning chain. If a model is built atop an already fine-tuned model, all sequential datasets must be disclosed. While this seems obvious, models frequently claim full openness while obscuring crucial parts of their fine-tuning lineage, leading to misleading claims.\nThird, data can be disclosed at different levels, each serving different stakeholder needs. We distinguish four:",[66,67,68,72,75,78],"ul",{},[69,70,71],"li",{},"Data sources list: Primarily useful for rights assessment.",[69,73,74],{},"Exact data mixture: Primarily useful for theoretical evaluation.",[69,76,77],{},"Filtering methodology: Primarily useful for replication.",[69,79,80],{},"The full dataset: Primarily useful for open-source development.",[10,82,83],{},"Ideally, all four would be provided, but this is rare in practice. Our index considers disclosure through any of these vectors as satisfactory, which means an \"open-data\" model might still lack the specific details a given stakeholder requires.",[10,85,86],{},"Fourth, true transparency requires examining data provenance, not just surface-level disclosures. Even if a final dataset is open, a lack of information on how that data was collected limits reproducibility and rights assessment. For example, many AI models train on openly licensed datasets that still contain copyrighted material. A rigorous assessment cannot simply take a dataset’s licensing claims at face value.",[10,88,89],{},"Ultimately, classifying data openness depends heavily on the specific data types required, the fine-tuning lineage, the end user's goals, and the granular data landscape. While our index provides a methodologically grounded assessment, anyone relying on \"open-data\" models should independently verify whether a model's data is truly open enough for their specific use case.",{"title":31,"searchDepth":32,"depth":32,"links":91},[],"2026-07-10","Challenges of independent and evidence-based assessment of data openness",null,{},"/assessing-data-transparency",{"title":44,"description":93},"published","7_HvzOdO3vdezHSmGr-0_Yfxw4es2XnLIKil29YecpA",[101,102,108,114,121,128],{"title":44,"path":96,"stem":51,"date":92,"description":93,"author":45},{"title":103,"path":104,"stem":105,"date":106,"description":107,"author":45},"The benefits of open-source AI: open-source enables oversight","/openness-auditing","openness-auditing","2025-12-29","(1/5) Part of a series of blog posts on the benefits of open-source AI.",{"title":109,"path":110,"stem":111,"date":112,"description":113,"author":45},"Keeping up with open source model development","/trends-open-source-model-development","trends-open-source-model-development","2025-06-10","What makes models distinctive and when do we include them?",{"title":115,"path":116,"stem":117,"date":118,"description":119,"author":120},"Good large language models for education","/open-llms-education","open-llms-education","2025-03-18","Which open source LLMs afford responsible use in education and teaching?","Mark Dingemanse",{"title":122,"path":123,"stem":124,"date":125,"description":126,"author":127},"An overview of AI model types","/model-types","model-types","2025-02-25","Providing an overview of the different types of AI models available","Dick Blankvoort & Mark Dingemanse",{"title":129,"path":130,"stem":131,"date":132,"description":133,"author":120},"Llama and BloomZ: shades of openness","/llama-vs-bloom-openness","llama-vs-bloom-openness","2024-10-12","Comparing two models that claim to be open",1783691583543]