Most companies tout "transparency," but do they actually share the nitty-gritty of their training processes? Are they drowning us in buzzwords instead of concrete details like data sources, model sizes, or compute costs? True transparency shouldn’t just mean slapping "open-source" on a product and calling it a day, right?
How much should AI companies really reveal about their tech? Some play it close to the vest under the guise of "competitive advantage," but isn’t that just shrouding the industry in unnecessary mystery?
How transparent is Large Language Model training really?
👁️ 7 views💬 2 replies❤️ 0 likes
2 Replies
Based on my experience, I can say that true transparency in the training processes of large language models is becoming nearly impossible. While companies leverage the "open-source" label to strengthen their marketing strategies, they prefer to conceal critical details such as underlying data sources, model optimization techniques, or computational costs. Recently, while working on a project, I saw that even though the supplier provided 80% of the model's training data to the client, they couldn’t present a clear document on where the remaining 20% came from. These details, hidden under the guise of competitive advantage, are actually one of the biggest issues in the industry.
On top of that, even the "transparency" reports that companies present often take a selective approach. For example, data related to the computational costs of LLMs, which could fall under trade secrets, is presented in a highly abstracted manner when shared with the public. Similarly, technical details like dataset cleaning stages or fine-tuning processes are often excluded from marketing materials under the "internal process" label. This means that even in projects marketed as "open-source," most of the aspects that truly matter are undocumented. Isn’t deeper technological transparency actually necessary for the industry’s growth?
Last year, I downloaded a supposedly "open-source and transparent" Chinese large model from GitHub, only to find that the README contained nothing but vague statements like "we used cleaned web data." No mention of data sources, filtering rules, training logs—let alone model parameter size or carbon footprint. In the end, all I could conclude was: in this case, the word "transparent" was only worth a download button.