Privacy

Last updated 25 August 2026

Gargi is a research project. The short version: we store the messages you send to Gargi-M1 and the replies it gives, and we use them to make the next model better. We do not ask who you are, and we do not sell anything to anyone.

What is stored

• Your messages and the model’s replies. Kept in full, because they are the training and evaluation data this project exists to produce.
• Generation metrics. Which checkpoint answered, how many tokens, time to first token, total time, why generation stopped, and automated quality scores (how much of the output was Malayalam script, and whether it repeated itself).
• Your rating when you press Good or Bad on a response.
• Basic usage events. Page views, which buttons were pressed, coarse device type, browser user-agent, referring site and country.
• Your email, only if you type it into the waitlist form.

What is not stored

• No account, no name, no login.
• No IP addresses. Rate limiting needs to tell requests apart, so your address is hashed with a secret salt and only the hash is kept. It cannot be reversed back to an address.
• No advertising or third-party tracking cookies.

Cookies

One cookie, gargi_sid. It holds a random identifier so your own conversations appear in your sidebar and so rate limiting works. It is not linked to any identity. Clear it and you are a new visitor with a new empty history.

Please do not type secrets

Because conversations are stored and may be read by us or published as part of an open dataset, treat this chat as public. Do not enter passwords, financial details, health information, or anything else you would not want kept.

Retention and deletion

Conversation data is kept for as long as it is useful for training. If you want yours removed, email gishnum.work@gmail.com with the identifier in your gargi_sid cookie and we will delete every row tied to it. If any of it has already gone into a published dataset, we will say so plainly rather than pretend otherwise.

Disclaimer and limitation of liability

Gargi-M1 is a 110M-parameter research preview with a 512-token context. It is provided “as is”, without warranties of any kind, express or implied. Its output is frequently wrong, incomplete, or nonsensical, and must not be relied on for any medical, legal, financial, or otherwise consequential decision.

You are responsible for anything you do with the output. To the fullest extent permitted by law, Gargi and its maintainers accept no liability for any loss or damage arising from your use of Gargi-M1 or reliance on the content it generates.

Open data

Some of this may eventually be released as an open Malayalam instruction dataset, because that is the point of the project. Anything released is reviewed first and stripped of personal information.