An NSFW filter for Marginalia search
marginalia.nu
March 30, 2026
13 min read
🔥🔥🔥🔥🔥
50/100
Summary
An NSFW filter for Marginalia Search is being developed in response to requests from API consumers. The filter will utilize a single hidden layer neural network, having explored various other methods prior to this decision.
Key Takeaways
- An NSFW filter for Marginalia Search is being developed using a single hidden layer neural network approach, focusing on speed and efficiency for CPU performance.
- The training of the classifier requires tens of thousands of samples, which are gathered by searching for NSFW content through the search engine.
- Open source models like ollama and qwen3.5 are utilized to label sample data as SAFE or NSFW, allowing for a more efficient training process without human intervention.
- Initial results from the classifier were poor due to a skewed training dataset, as most samples were collected from NSFW-related queries, affecting the model's accuracy.