Training data simply does not help you here. Our existing architectures are not explainable or auditable in any meaningful way, training data or no training data.
I don't necessarily agree and suggest the Open Source Definition could be extended to cover data in general (media, databases, and yes, models) with a single sentence, but the lowest risk option is to not touch something that has worked well for a quarter century.
The community is starting to regroup and discuss possible next steps over at https://discuss.opensourcedefinition.org
To the extent the problem is intractable, I think kt mostly reflects that LLMs have an enormous amount of training data and do an enormous amount of things. But for a given specific problem the training data can tell you a lot:
- whether there is test contamination with respect to LLM benchmarks or other assessments of performance
- whether there's any CSAM, racist rants, or other things you don't want
- whether LLM weakness in a certain domain is due to an absence of data or if there's a more serious issue
- whether LLM strength in a domain is due to unusually large amounts of synthetic training data and hence might not generalize very reliably in production (this is distinct from test contamination - it is issues like "the LLM is great at multiplication until you get to 8 digits, and after 12 digits it's useless")
- investigating oddness like that LeetMagikarp (or whatever) glitch in ChatGPT