Let's discuss how much transparency is needed in open-source models. Options:
- Fully open (model components, training data, weights)
- Only weights with the implementation code
- Technical documentation with API access
What do you think? Which one would you prefer and why?
What should transparency in LLMs look like: Just the source code?
👁️ 4 views💬 2 replies❤️ 0 likes
2 Replies
Last year, I was working on a cloud-native LLM project for a fintech client that required strict compliance. We started with option 2—just the weights and code—but hit a snag when auditors asked for training data lineage. We couldn’t trace why certain biases crept in because the data pipeline wasn’t documented.
So we pivoted to option 1: full transparency, including the data preprocessing scripts. It took us three extra sprints, but when regulators demanded an explainability report, we basically handed them the Jupyter notebooks with notes from every epoch. Saved us weeks of back-and-forth.
That experience skewed me toward “more open than less.” Losing a little competitive edge beats losing a banking license because you can’t prove your model didn’t cheat on fairness metrics.
When I say "open" while opening, I imagine it 🤣 with the hope of clarity, I say "publish the data" as a beginner who couldn't even learn Flask in 2 months. The best option is the third one: sharing my past with API access along with "I was a beginner too" 😅