internet-explorer-ssl/internet-explorer

Iteratively cycles through four steps: querying the Internet for images via text search, self-supervised training on downloaded images, relevance scoring against the target dataset, and query refinement for the next iteration. Targets vision model fine-tuning tasks where task-specific data is scarce, demonstrating performance matching CLIP oracles on multiple datasets using only single-GPU resources over 30–40 hours of web exploration.

163 stars. No commits in the last 6 months.