[FAQ] Kokoro may have been review bombed on Dec 29
There is currently a large discrepancy between the Arena snapshot and the current Elo rating of Kokoro listed on Pendrokar's TTS Spaces Arena. Some have fairly asked why this is the case—skepticism is good given what happened with Reflection 70B.
Please refer to this discussion in TTS_Arena: Unusual voting pattern detected (Dec 29th). Props to Pendrokar for spotting this so quickly.
Although I personally agree that those votes are likely not legitimate, I defer entirely to @Pendrokar on arena policy. It could set a bad precedent if model makers point at Arena votes that they don't like and demand their removal.
Irrespective of the decision to remove or keep, there are ideas to defend smaller arenas against such attacks (including artificial boosting) like time-based smoothing (limit the contribution of any particular day's worth of votes to Elo rating), or authenticated votes (sign in with HF button). However, the former could be mathematically complex to implement or favor incumbents, and the latter could reduce total vote counts & statistical significance across the board (unless it is added as an option, not a requirement).
Also, I would discourage people from doing any shenanigans in the other direction either. Generally we should strive to keep Arena votes anywhere as unbiased as possible.
Edits: Phrasing, also editing with stale browser cache is causing consecutive edits to flip-flop between "legitimate" and "illegitimate"
Sorry this happened to you. It seems quite obvious those votes were tampered with. I guess this is one of those: "this is why we can't have nice things". A technical solution should be viable though, where the URLs for the samples are randomized and "decrypted" server-side. From the generation code I've seen from you, it also seems like there should be many samples, but when I went on there to rate a few models, I pretty much got only ~2-3 samples for every model.
Oh, and by the way. Thanks a lot for opening up your model! I like it quite a lot, especially for it's size.