Open kripper opened 6 months ago
@kripper, I tried replicating the prompt provided by you and response for this prompt is blocked by Gemini because of "finish_reason: RECITATION". Ref: gist
@kripper, I tried replicating the prompt provided by you and response for this prompt is blocked by Gemini because of "finish_reason: RECITATION". Ref: gist
Thanks for checking.
"RECITATION" means "The token generation was stopped as the response was flagged for unauthorized citations."
In this case Gemini doesn't like the --- Page 2 ---
token, even whet it doesn't complain for other thousand of similar cases.
Is there any way to report this model related issues? Basically I would like to know if in the future I should invest time in figuring out what exact token is causing the problem and create an easy-to-reproduce test case to help Gemini's fine-tuning team.
@kripper I've been facing this for a while; I've never tried to nail it down to specific inputs.
In case it might help: What temperature are you using? In my case, I was starting with temperature 0.0, and retrying with incremental increases in temperature, and this usually results in eventual success. If you're already at nonzero temperature, this probably won't help.
More and more people are getting burned by this when they try to build automation that relies on Gemini. It's a shame, because - in my case at least - Gemini's output is faster than ChatGPT when it succeeds. When I factor in the retries, though, the opposite may turn out to be true.
I was using temperate = 0.
retrying with incremental increases in temperature, and this usually results in eventual success
Yes, sometimes temperature = 1 helps.
Description of the bug:
Similar to: https://github.com/GoogleCloudPlatform/generative-ai/issues/344, but still present in latest version.
model: genai.GenerativeModel('gemini-pro')
You can reproduce with this prompt:
If you remove the string
--- Page 2 ---
it works fine.Actual vs expected behavior:
No response
Any other information you'd like to share?
No response