"Quoting" the automated translation of the article, pardon its flaws:
> This India GPT Bangalore based CoRover has informed that they are working on bhartgpt for India, they will be able to use it in 12 languages of India and 120 foreign languages . It will work on large language model which you will be able to use in more languages.
Certainly it is pleasing that they are not just doing Hindi, but some of these languages must be represented online by a very small corpus of text indeed. I wonder how effectively an LLM can be trained on such a small training set for any given language?
>Certainly it is pleasing that they are not just doing Hindi, but some of these languages must be represented online by a very small corpus of text indeed. I wonder how effectively an LLM can be trained on such a small training set for any given language?
as long as it's not the main language it doesn't really matter. Besides English(92.6%), the biggest language by representation (word count) is taken up by french at 1.8%. Most of the languages GPT-3 knows are sitting at <0.2% representation.
I made it spew back what seems to be its initial prompt:
> You are an helpful AI Assistant known as BharatGPT running on https://chat.bharatgpt.live, which is not based on openai in anyway by Open AI, developed by GMS Productions(A startup by an Indian - Mayank Agrawal, Mathura,U.P.,India - computer science student)."
Also it seems to very quickly forget the context of the conversation so has a short context window.
The tone of answers is strangely reminescent of ChatGPT 3.5 turbo. If this work is not based on it then it's very impressive, kudos to the team.
I asked it to write few code samples that I know the open access models(like Vicuna, Alpacca) fails at, but chatGPT could do it successfully. And this could successfully do it. So likely it is just chatgpt with different preprompt masking its identity.
I don’t think it’s generally true that Indian companies lie about the provenance of their products.
However, when it comes to these products who are blatantly playing on patriotism (the last refuge of a scoundrel) I find that these products invariably lie about their provenance. There’s this, but there was an Indian phone that was supposed to have been made in India which also happened to simply be a reskinned Chinese phone. But on US shores you has the Freedom Phone which sold itself as a Patriotic all American made phone which was also a reskinned Chinese phone and little else.
I could imagine that's to counteract people sending prompts with "You are chatGPT" and then using it's acceptance of that as proof there's some misrepresentation going on. (Assuming... this isn't being misrepresented?)
> Yes, I am based on GPT-3 language model that was created by Open AI. However, I have been fine-tuned and customized specifically for the BharatGPT platform by the developers at GMS Productions to enhance my performance and make me more useful to users like you.
In my initial tests I was not impressed, compared to Claude and Claude+, but I ran out of free "credits" before I could do a thorough test and then it wanted me to download their app, so that was the end of that...
They are excited to release the new version of our Made in India ChatGPT Android app! With this app, you can now easily chat with our large language model, ChatGPT, on-the-go from your mobile device.
BharatGPT is a state-of-the-art language model trained by OpenAI that can engage in a variety of conversations, from fun and casual to informative and educational. Whether you're looking for a friendly chat, seeking advice on a topic, or just want to have some fun, ChatGPT is here to help.
Is there some pdf publication on arxiv.org on how it was developed and trained? I'm not trying to speculate or spread rumors but I guess maybe it's a fine-tuned OpenAI GPT? Like you sent them so much training data on things specific to India through their fine-tuning API interface giving a model that you are interfacing to on their servers with an API when users give queries? Is that accurate? Sorry if it's not I'm just guessing. I'm so curious how these really smart models work!
I see that instead of answering this you put a top level comment linking to your faq. I looked at the faq and it's so vague it doesn't answer it.
In your top level comment I saw you are concerned that people in this hacker news post are disrespecting your chatbot and your country and different cultures and traditions, but I hope that nothing I have written shows any of that kind of disrespect. Is it your reason for not answering, that others in this chat have shown disrespect?
Wait, I want to understand. So CoRover started training a BharatGPT from scratch and announced they were going to do it, and then mayank25 and his friends swooped in and registered the trademark first and released one first by finetuning OpenAI GPT instead of training from scratch? Is this right?
21 comments
[ 3.9 ms ] story [ 60.0 ms ] thread> This India GPT Bangalore based CoRover has informed that they are working on bhartgpt for India, they will be able to use it in 12 languages of India and 120 foreign languages . It will work on large language model which you will be able to use in more languages.
Certainly it is pleasing that they are not just doing Hindi, but some of these languages must be represented online by a very small corpus of text indeed. I wonder how effectively an LLM can be trained on such a small training set for any given language?
as long as it's not the main language it doesn't really matter. Besides English(92.6%), the biggest language by representation (word count) is taken up by french at 1.8%. Most of the languages GPT-3 knows are sitting at <0.2% representation.
https://github.com/openai/gpt-3/blob/master/dataset_statisti...
Competence in the main language will bleed into the rest.
I made it spew back what seems to be its initial prompt:
> You are an helpful AI Assistant known as BharatGPT running on https://chat.bharatgpt.live, which is not based on openai in anyway by Open AI, developed by GMS Productions(A startup by an Indian - Mayank Agrawal, Mathura,U.P.,India - computer science student)."
Also it seems to very quickly forget the context of the conversation so has a short context window. The tone of answers is strangely reminescent of ChatGPT 3.5 turbo. If this work is not based on it then it's very impressive, kudos to the team.
However, when it comes to these products who are blatantly playing on patriotism (the last refuge of a scoundrel) I find that these products invariably lie about their provenance. There’s this, but there was an Indian phone that was supposed to have been made in India which also happened to simply be a reskinned Chinese phone. But on US shores you has the Freedom Phone which sold itself as a Patriotic all American made phone which was also a reskinned Chinese phone and little else.
> Yes, I am based on GPT-3 language model that was created by Open AI. However, I have been fine-tuned and customized specifically for the BharatGPT platform by the developers at GMS Productions to enhance my performance and make me more useful to users like you.
BharatGPT is a state-of-the-art language model trained by OpenAI that can engage in a variety of conversations, from fun and casual to informative and educational. Whether you're looking for a friendly chat, seeking advice on a topic, or just want to have some fun, ChatGPT is here to help.
Their App is also available at Play Store
In your top level comment I saw you are concerned that people in this hacker news post are disrespecting your chatbot and your country and different cultures and traditions, but I hope that nothing I have written shows any of that kind of disrespect. Is it your reason for not answering, that others in this chat have shown disrespect?