The Rise of Indian Llamas
Llama 2 has been one of the most downloaded open-source language models globally, with close to 757K downloads on HuggingFace. In India, we are witnessing a new wave of Llama 2 adoption, where users are building ‘Local Llamas’ to support low-resource regional languages in the country, besides burgeoning enterprise adoption. This is just the beginning.
Sarvam AI recently introduced the first version of OpenHathi, marking the release of India’s first Hindi LLM in the OpenHathi series of models. Built on a cost-effective platform, this model, extending from Llama 2 7B, exhibits performance similar to that of GPT-3.5 for Indic languages.?
Then there is Tamil Llama, developed by Kaggle Master Abhinand Balachandra. This Indic LLM is engineered to elevate the Tamil language domain and has been built on top of Meta’s Llama 2.?
Another avid developer based out of Hyderabad, Ramsri Gourham Golla, is working on building a Telugu Llama. Meticulously trained on Llama 2, his LLM showcases remarkable efficiency in the token count for Telugu text. As per Golla, the Telugu language consumes fewer tokens as compared to English.?
A group of developers from Odisha are working on building Odia Llama, aka OdiaGenAI, a Llama 2 finetuned LLM for the low-resource Odia language using a prepared instruction set containing domain knowledge.?
Not just in India. There has been a significant rise in the number of vernacular models globally built on top of Llama 2. Some of them include SeaLLMs, VinaLLaMA, LLaMAntino, and Dutch Llama among others.?
Check out the complete story here.
AI Forum for India. Our Discord Community for AI Ecosystem, In collaboration with NVIDIA.?Join Today >>
Top Stories of the Week >>?
AIM Top Ranked PG Data Science Programs (Online/Hybrid) – 2023
After our?Full Time On-Campus programs rankings, we are our with our postgraduate program ranking for the year. Access the complete ranking here.
Should India Regulate Foundational Models Akin to the EU?
There is no doubt that the EU has set a precedent by becoming the first jurisdiction in the world. India is now closely examining the act and how it develops, but the question that arises is whether it should take a similar approach. Read the full story to find out.?
The Ultimate Satellite Internet Showdown: Kuiper vs Starlink
When Amazon chief Andy Jassy shared a post about Project Kuiper broadband satellites on LinkedIn, it looked like SpaceX’s Starlink on steroids. What’s even more interesting is that SpaceX is delivering its competitors’ satellites into space. But why? Read to find out.?
Apple Smoothly Crafts ‘Mouse Traps’ for Humans
Apple surely knows how to develop amazing products and thrives on its ever-evolving ecosystem. But, what many don’t know is, it’s a trap – the ‘Apple Trap’ – for once you’re in, you are a happy captive. Read the full story to know more about how Apple plays its tricks and treats here.
Synaptics is Open to Packaging its Chips in India
领英推荐
In an exclusive interaction with AIM, Michael Hurlston, CEO of Synaptics, emphasises the company's unique approach in India, focusing on end-to-end semiconductor design, testing, and implementation, while expanding into IoT, AI, and wireless connectivity.?
“If you look at the semiconductor companies like Qualcomm, AMD, or Intel, which have a footprint in India, what they typically do is they do a piece of the solution here and they do a piece of the solution in the US and they do a piece of the solution in Europe,” said Hurlston, stating their ambitious plan to build a comprehensive solution locally in India.
Check out the complete interview here.
AIM VIDEOS >>
[Reaction Video] India’s LLM, Ola Krutrim is Here ft. Vandana Nair
Ola's Krutrim still feels like it's too good to be true. In our latest episode of the reaction video, AIM dives deep into decoding some of the specific claims and announcements made at the recent Krutrim AI launch event by Ola’s chief Bhavish Aggarwal.
AIM EVENT >>
February 1 to 2, 2024 | NIMHANS Convention Center, Bangalore?
THE BIGGEST GENERATIVE AI CONFERENCE IN INDIA?
6th Edition 1,500+ Attendees 50+ Speakers 400+ Organizations
Regular Passes to expire on 12th Jan. Book your passes now >>
AIM SHOTS >>
Bhaiyabot's new AI model, rohansolo/BB-L-01-7B, surpasses Sarvam AI's OpenHathi-7B, Meta Llama-7B, and Mistral AI's models in benchmarks, maintaining its base model's performance. Read more here.?
Tech Mahindra’s former chief CP Gurnani recently shared the development of an LLM specifically designed for Hindi and its 37 dialects. Check it out here.?
Anthropic are in advanced talks to secure $750 million from Malo Ventures in funding to bolster the development of a safety-focused conversational AI chatbot, outshining OpenAI.?
Google researchers recently introduced VideoPoet, a new large language model (LLM) for video generation. It looks better than RunwayML, Stable Video Diffusion, Pika, and the most recent ‘Animate Anyone.’
Apple recently announced that it has tackled the significant challenge of running large language models (LLMs) on devices with constrained memory capacities. Read more here.
???????????? ???????? ????, ????????, 7 ??????????????, ???????????????????? ????, ????????????, ?????????????? ????????????????????, ??????????????, ????????????. ?????????? ?????? ????????????????
11 个月Llama 2 is not very stable and tends to hallucinations a lot both in generative as well as in RAG in my experience. Any tuned version of Llama 2 would exhibit foundation model characteristics. When it is done for additional languages, there would be additional issues as most of these models are tuned by adding more layers. No one is reporting if there is any loss of previous learning.