Every time I try to get Gemini to generate Latin, it stops after only a few tokens. I believe this is due to self-censoring?
1 Like
Welcome @monsterslayer
In case your outputs are getting cutoff mid-sentence you can reduce the the safety settings under advanced settings.
2 Likes
Welcome in!
As previously suggested, the default safety settings can cut off responses if the response is flagged at medium to high risk. Turning those dials down should allow it to pass through.
What, specifically, flags content as being of risk we donât yet know. The flags seem to be an active work in progress.
Otherwise, I tried a single shot prompt myself with the default safety settings, and it generated fine.
At this point, this might just take some trial and error as the kinks get worked out.
2 Likes
I can confirm the original finding, there are difficulties with Latin that arenât happening with Athenian Greek.
Example chat session here : https://aistudio.google.com/app/prompts?state={"ids":["1fc97RsCUzhK94JiP4pgDMZe1eAe2MJqj"],"action":"open","userId":"114091953393014428154","resourceKeys":{}}&usp=sharing
It chokes after starting out just fine. Itâs not the safety settings in AI Studio, it was indicating âHarassment Lowâ and my settings were higher.
Asking it to continue the dialog between the senators doesnât help either. It starts over and chokes again.
I think the time travel agency will have to scrap the travel plans for Romeđ
2 Likes
Update - itâs a different kind of censorship. I removed the young senator from the setting and requested the dialog in Latin again. It chokes and the message is
Not something one can control through the safety settings.
2 Likes
Additionally validated through direct API call (code, not AI Studio). The server response to an attempt to get Latin is alternately HTTP status code 500 or
{
"candidates": [
{
"finishReason": "OTHER",
"index": 0
}
],
"usageMetadata": {
"promptTokenCount": 586,
"totalTokenCount": 586
}
}
2 Likes
Tested with the latest released version, hoping that GA would solve this odd behavior. @keertk-google , is there any reason why the output from the Gemini (both 1.5 flash and pro) is blocked with âOTHERâ when they translate English into Latin, while having no difficulty translating Latin into English?
It looks like Latin is getting unfairly discriminated against. As the original poster mentioned, Ancient Attic Greek works perfectly well both ways.
2 Likes
I guess it stops the response when a subsystem determines a language used is in blocklist. I hope Google doesnât make any such limitations. They are doing so because of load over their servers, and to reduce them
Yes, we are certain it is a separate subsystem (a gateway) in front of the Gemini server, because the block is inserted after streaming had started. A few hundred bytes get through and then this gateway system clamps down and shuts the Gemini server up.
Whether that has anything to do with data center load overall is pure speculation. Personally, I think it is simply poor implementation of the block criteria that affect how the API is perceived by the outside world in a negative way. Thereâs both the inexplicable cases of finishReason OTHER like what does Google have against Latin and the unexpected and problematic behavior for finishReason MAX_TOKENS.
2 Likes
Maybe, they are uncertain about the response and due to criticism, they have censored it? 
1 Like
Update: the inexplicable block for Latin in Gemini 1.5 has been removed. The translation into Latin is pretty good (actually, I believe it always would have been, based on the little amount that previously could get through before getting blocked with finishReason set to OTHER).
This might be a good opportunity to test whether Indian subcontinent regional languages are now also allowed (I am not capable of trying such a test).
2 Likes
It has been a time I have been into this forum.
Yeah, it has been fixed and doing well⌠Great thanks to Google for fixing it. Also thanks for the community especially @OrangiaNebula to contributing in bringing attention to it.
thank you all