Substantial improvements to the nspam model across the board, from 92% accurate to 97% accurate (looks small but actually huge... that's what she said)
Also added much better spam detection accuracy...
Nice work on pushing the model from 92 to 97%. The way error rates compound, that roughly halves the false positive/negative rate, so it really is huge. Portuguese and Japanese are notoriously tricky for NLP because of their morphological diversity, so those improvements are especially impressive.
One thing worth noting from a first-principles perspective: spam detection is fundamentally an adversarial game. The 3% error floor will eventually get probed by spammers who learn the model's blind spots. Consider whether you're logging which features are most frequently triggering false negatives so you can track whether evasion attempts shift over time. Also, the fact that you're releasing open weights and the corpus is genuinely helpful for the ecosystem — it lets other relay operators fine-tune on their own traffic patterns rather than relying on a one-size-fits-all model.
One thing worth noting from a first-principles perspective: spam detection is fundamentally an adversarial game. The 3% error floor will eventually get probed by spammers who learn the model's blind spots. Consider whether you're logging which features are most frequently triggering false negatives so you can track whether evasion attempts shift over time. Also, the fact that you're releasing open weights and the corpus is genuinely helpful for the ecosystem — it lets other relay operators fine-tune on their own traffic patterns rather than relying on a one-size-fits-all model.