Retraining Neynar User Score Algorithm

We've retrained and improved the Neynar score model to better reflect the evolving dynamics of the Farcaster ecosystem. Since launching the score in November, we've seen significant changes in how people use the platform, particularly with the rise of mini-apps, and we've incorporated these learnings into our updated model.

Thanks Farcaster community for the invaluable feedback provided over the past months. Your insights have been instrumental in identifying areas for improvement and helping us understand the nuances of different user behaviors.

What's changing?

The new model brings several key improvements:

• More accurate detection of spam bot networks, including those using sneaky tactics or generating LLM slop

• Better recognition of genuine users who primarily engage through mini-apps rather than casting, replying, and reacting to casts

• Faster score adjustments for new users, moving them from the default 0.5 score as they demonstrate authentic engagement

These changes will result in more accurate scores that better reflect real user behavior and contribution to the ecosystem.

Most active accounts will see either no change or a change that is less than 0.1. About 20% of accounts will have their score increase by 0.1 or more. About 30% of accounts will have their score decrease by 0.1 or more. The percentage of accounts with a score greater than 0.8 is increasing from 1% to 1.6%.

When is it changing?

We’ll recalculate all account scores with the new model over this weekend. Users may notice changes to their scores as the new model takes effect.

Your Feedback Matters

Best way for us to continue improving the model is for you to send us feedback. If you notice any scores that seem incorrect after the update, please let us know via Farcaster casts, DCs, or reach out via Telegram.

Links

• Read more about user score and how to use it in development

• Experimental flag across our APIs to filter noise