World's first AGI in Sanskrit

[prabhat kumar singh]

Namo Namah.

I am writing after 6 months of last announcing the success achieved in creating skip-thought vectors of Sanskrit and very high BLEU score in NMT. I am releasing for public preview using web interface. It is tensorflow based in backend and looking to collaborate. It is world's first AGI system(Artificial General Intelligence). Due to many security and ethical concerns, I am hesitant to make it public, so also soliciting ideas for how to scale this work.

Here is the blog detailing the current state and way-forward.

https://अथ.भारत/

https://medium.com/@prabhatkr/%E0%A4%85%E0%A4%A5-%E0%A4%AD%E0%A4%BE%E0%A4%B0%E0%A4%A4-sanskrit-framework-for-agi-artificial-general-intelligence-82ec6b0c2d2

Please feedback & critique,

Namo Namah.

Prabhat

[विश्वासो वासुकिजः (Vishvas Vasuki)]

What can I use this website for? Could you demonstrate its use?

If I type a sanskrit sentence, I get:

खेद, Error: 102e

​Also, could you elaborate on the " many security and ethical concerns​" you refer to?

--
You received this message because you are subscribed to the Google Groups "sanskrit-programmers" group.
To unsubscribe from this group and stop receiving emails from it, send an email to sanskrit-programmers+unsub...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

--

--
Vishvas /विश्वासः

[prabhat kumar singh]

Namo Namah!

Apologies for delay in responding.

I didn't expect 50 parallel users and it had crashed. So was trying to put a load balancer.

In terms of use, if you enter an english text, the system will translate the text into sanskrit.

Since it is machine model, it has learnt the meanings from Valmiki Ramayan which I got from IIT Kanpur(with permission).

I am working to extend it into more AGI centric system: Learn One , Improve All.

Please feedback. Sorry again for inconvenience.

Thanks

[Dhaval Patel]

I tried three four sentences of two or three word length. Results were not at all related.

Ram loves Sita

छन्ना भव

You look beautiful

तां विद्धि

They lived in forest

विश्वामित्रस्सु

I guess there was something broken in code or was it supposed to send such garbage output only?

[prabhat kumar singh]

Thanks Dhaval,

I am glad you tested this. You have fame. Many people recommended to reach you for curated datasets though it didn't happen before.

A few things are lacking.2 major are:

1. Training iterations: Currently the model has achieved 100K iterations. But a minimum of 5Million iterations are required for accuracy.
2. Data set: The training is done on ramayan sourced from IITK, which is only good sentence to sentence marked up translation. I hope that accuracy will increase. But more of effort has to be put on creating better dataset for training general machine models. There are many benefits of having done so.

To infer the result, I see that the model has identified the context of forest with Vishwamitra. But yes, it won't be so apparent , thus looks garbage.

It has only learnt less than 0.02% (=100K of 5 Million) currently.

It is very costly to train models, so I have setup my resources at home for best optimal output.

I liked your feedback. Please keep them coming.

Best regards,

Namo Namah

[ken p]

Somehow translator doesn't work upon pasting a simple Sanskrit sentence.

खेद, Error: 102e

https://xn--l1b8d.xn--h2brj9c/

[prabhat kumar singh]

This might be due to cache of an older version. I have published some changes yesterday.

Please try after clearing the cache.

I tested, and works fine.

hi there

तो महा

The system is under learning iterations, currently at 100K+ (0.02% complete), so it is acting as a context translator. 

Namo Namah,

Best Regards