#OctoAI should be marked as a provider that logs requests

7 messages · Page 1 of 1 (latest)

mighty solar
#

In OctoAI's documentation they explicitly state:
Customer prompt data is persisted in logs for 15 days and then deleted.
Which to me defintively qualifies as logging requests.

Also I'd argue that DeepInfra should potentially also be marked as such given that they state:
"We reserve the right to look at and log a small portions of requests when necessary for debugging or security purposes."

Which effectively means that any random prompt you send in could be inspected by a human.

Also I'm not sure the help section is the right place to report issues like this, but it seemed like the most appropriate section. Please inform me if there is a better place for this type of inquiry.

turbid stream
#

The mark is primarily used to indicate that the provider can 'train' their own models based on your prompts.
Unlike some other domains like VPNs, for LLMs, we're not aware of any provider that claims to log absolutely nothing about the request while maintaining their service, delivering production-level performance, and undergoing third-party audits.

mighty solar
#

No, I'm fairly certain that is incorrect. The icon it is used for both, but the tooltip that comes up when you hover over it states whether the provider just logs requests or logs and trains on requests. There is also a setting in OpenRouter's account panel that disables hosts which train on your data, which does not disable most providers with this icon.

For instance if you look at the icon for NovitaAI (which is still enabled if you disable data training) it only states:
Your requests may be logged by this third-party model provider.

Whereas the icon for Mancer (which is disabled for data training) it states:
Your requests may be logged by this third-party model provider. They may also use your prompts and completions to train new models.

And having looked at all of the icons it appears that most of the providers with that icon is not listed as training on the logged requests. So I'd disagree that it's primary purpose is to indicate such providers.

Also there are providers that state or provide options to not log any prompts at all. Which includes fireworks.ai and together.ai. While it's true that you have to take them on their word, as I'm fairly certain none of them has been audited. They at least claim to not log requests, which is worth something.

turbid stream
#

Ah you meant the logging for prompts. Yes you are correct.

#

There are some unintuitive aspects of our UI.
The mark can be displayed and can be filtered accordingly 'when we know our provider's written policies and the corresponding URL.'
When there is no mark, we do not guarantee that the provider is not logging.
It's usually manually maintained, and for the providers you've mentioned, we'll update with their latest policies.

#

Thanks for spotting :>

turbid stream