Рубрика: AI

Founded in 2015 by Nandan Nilekani, Rohini Nilekani, and Shankar Maruwada, EkStep Foundation focuses on building digital public goods and delivering large-scale impact for billions of people.

Working towards increasing access to learning opportunities for millions of Indian children, the company has developed Sunbird, a digital public good which powers DIKSHA, a national platform for school education led by the Ministry of Education in India.

However, now the foundation wants to leverage AI for the betterment of Indian citizens.

In June last year, the People+AI initiative was born out of the EkStep Foundation. It’s a community-driven effort that aims to harness AI capabilities to help a billion people. It brings together a diverse group of individuals, innovators, doers and tinkerers to find, test, and scale AI capabilities to address societal problems at scale.

Building AI use case repository

Tanuj Bhojwani, who heads the People+AI initiative, told AIM that the foundation had been exploring the benefits of AI for quite some time. However, it was only in the past year that they decided to establish a specialised unit focused on extensively exploring AI use cases.

He believes that in a decade, India is going to be the AI use case capital of the world because of its huge population, diversity, languages and the need for specialised AI tools, be it in healthcare, education or any other field.

“However, the biggest stumbling block right now is the lack of understanding regarding the potential use cases of AI. While building chatbots is undoubtedly crucial, there is a significant amount of work that still needs to be undertaken, especially in the context of India,” Bhojwani said.

Hence, People+AI is building a use case repository, which is an ongoing project focused on creating a structured library for AI use cases. This initiative involves defining a contribution model and designing a user experience to actively encourage and assist contributors.

“Our goal is to document and catalogue AI use cases that are particularly relevant and unique to India. Emphasis will be placed on identifying cases that are distinctive and specifically applicable to the Indian context,” Bhojwani added.

Building AI language tools

India is known for its linguistic diversity with over a thousand different languages and dialects spoken across the continent. EkStep Foundation has funded AI4 Bharat, an initiative led by the Indian Institute of Technology (IIT) Madras, focused on building open-source language AI solutions for Indian languages.

AI4Bharat is actively creating datasets for various Indic languages to train AI models. In many cases, the data collection process often requires visiting remote locations to record local dialects. To facilitate this, there is a requirement for various tools and resources.

People+AI is involved with stakeholders to determine if better tools and resources can be developed. “For this, we are collaborating and talking to different organisations, some in Africa, because they are also facing similar problems,” Bhojwani said.

A leaderboard for IndicLLMs

Moreover, People+AI is actively involved in developing a benchmarking leaderboard for Indic large language models (LLMs).

Several Indic LLMs and SLMs have emerged in the past few months, such as Tech Mahindra’s HindiLLM with 539 million parameters and Sarvam AI’s OpenHathi, a 7 billion parameters model trained on Hindi, English, and Hinglish.

However, there is no means to determine which of the two models mentioned above works better for certain use cases, say healthcare, for example.

“A Malayalam-speaking nurse is going to speak in English when referring to medical terms, because such terms may not be commonly expressed in Malayalam. In such scenarios, a Malayalam language model (LLM) might not be effective. Instead, a model trained on both Malayalam and English medical datasets would be more beneficial,” Bhojwani said.

The objective is to establish a leaderboard to determine the effectiveness of models for various Indian languages. The leaderboard will feature default rankings and provide an option for users to “upload their own data”.

Hence, there is a growing need for an Indian use case-focussed benchmark and for this, People+AI is talking to different stakeholders in the Indian AI ecosystem such as AI4Bharat, Indian Institute of Science (IISc) as well as folks from BharatGPT.

However, Bhojwani notes the project is still in its very initial stage.

Democratising AI computing in India

If India is going to be the AI use case capital of the world, a substantial compute infrastructure is imperative. People+AI is actively working on developing an open compute infrastructure network to meet the rising demand for compute while promoting market competitiveness.

“One of the ideas is can we decentralise computing infrastructure and create an open network where compute providers and customers who need this infrastructure can interact and discover each other,” Tanvi Lall, director of strategy at People+AI, told AIM.

The idea here, according to Lall, is to help a small business based in Tier 2 or Tier 3 cities access computing infrastructure for training or inferencing at a lower cost compared to leveraging services from the likes of AWS, Google or Azure, which could prove to be costly.

People+AI is already talking to Indian computing service providers like NeevCloud and Vigyan Labs and the idea is to create a network of micro-data centres with interoperable standards, allowing small businesses and startups to easily plug and play, based on their specific requirements.

The post This Nilekani-backed NPO Aims to Make India the Global AI Use Case Capital appeared first on Analytics India Magazine.

Joining the league of indic LLMs like Telugu, Malayalam, Tamil and Oriya Llama, is MahaMarathi 7B. Boasting 7 billion parameters, the Marathi LLM is built on Meta Llama-2 and Mistral AI framework.

With computing resources and data provided by Microsoft for Startups funded company CourtEasy.ai, this open-source LLM is domain-adapted, continually pre-trained, and instruction fine-tuned using the Meta Llama-2 and Mistral AI framework.

The brains behind this research include Dr. Aakash Patil, postdoctoral researcher at Stanford University, Mrunmayee Shende, cofounder of CourtEasy.ai, and Niraj Singh, ML engineer at Inbound Health.

To democratise ML research, the team has released the initial version of the pre-trained base model on Hugging Face, inviting developers, startups, and public and private organisations to innovate by developing fine-tuned models for various use cases.

Marathi’s unique linguistic characteristics, complexity, and cultural context are addressed by MahaMarathi 7B, making it suitable for handling complex conversations and instructions. The language model is available for free on Hugging Face, promoting broader access and encouraging applications in various fields, including business and e-governance.

Marathi, spoken by over 83 million people predominantly in Maharashtra, is the 13th most spoken language globally and the third most common in India. Acknowledging Maharashtra’s significant economic contribution, with Marathi businesses and consumers contributing over 15% to India’s GDP, the MahaMarathi 7B aims to catalyse innovation in the region. The creators envision the potential impact of this Marathi LLM on diverse sectors such as skill training, education, healthcare, agriculture, environment, urban planning, and traffic management.

The release of the Marathi LLM is a step towards making AI more accessible and applicable to non-English languages. The team plans to release instruction-tuned and preference-optimised models in the coming months.

The post Meet the New Indic LLM, MahaMarathi 7B appeared first on Analytics India Magazine.

Leading education company upGrad is harnessing AI to translate popular learning materials like bootcamps and certifications into Indian languages.

Presently, more than 40% of upGrad’s learners in India hail from Tier 2 cities and beyond, where local languages are prevalent. In response to this demographic insight, the company aims to translate its offerings into Hindi, Tamil, Telugu, Kannada, and Bengali in the initial phase.

To enhance accessibility and effectiveness, upGrad will integrate a blend of pedagogy and human interventions. Local language-speaking counsellors and buddies will be deployed to guide learners through the educational journey.

“Considering that only 7 to 8% of graduates in India are deemed employable, our objective is to substantially elevate this figure by providing millions with some initial support or springboard to get closer to their career aspirations,” said upGrad president Asheesh Sharma.

Sharma highlighted the company’s investment in AI, stating that they have been using AI in operations for a while and are now integrating it further into the curriculum and teaching methods. This strategic move aligns with the company’s growth and transformative goals for the fiscal year 2025.

The focus in Phase 1 will cover domains like engineering, data science, AI, cloud computing, DevOps, UI/UX, and agile project management, with Phase 2 expanding into areas such as cybersecurity, blockchain, product management, business analysis, and ITIL, along with foreign languages like Spanish and Chinese.

With over five lakh enrolled learners, the demand for upGrad’s courses has been notable in key cities such as Bengaluru, New Delhi, Mumbai, Hyderabad, Odisha, Pune, Chennai, and Kolkata. To address inquiries effectively, the company will leverage an in-house tech tool, aligning with the Government of India’s initiative to connect culturally and foster cognitive development among the youth.

The post upGrad is Using AI to Translate Learning Materials into Indian Languages appeared first on Analytics India Magazine.

Google has unveiled Gemma, a new family of open models, leveraging the research and technology behind the existing Gemini models. The Gemma open models are released in two sizes, Gemma 2B and Gemma 7B, each offering pre-trained and instruction-tuned variants.

Users can start working with Gemma today using free access in Kaggle and a free tier for Colab notebooks. Additionally, first-time Google Cloud users can avail themselves of $300 in credits. Researchers can also apply for Google Cloud credits of up to $500,000 to accelerate their projects.

Gemma outperforms Llama 2 on several benchmarks, including MMLU, HellaSwag, and HumanEval.

To support developer innovation and responsible use, Google is also providing a Responsible Generative AI Toolkit alongside the models. This toolkit includes essential tools for creating safer AI applications with Gemma, offering guidance and support for developers.

To facilitate widespread adoption, Gemma is compatible with major frameworks, including JAX, PyTorch, and TensorFlow through native Keras 3.0. The release includes ready-to-use Colab and Kaggle notebooks, integration with popular tools such as Hugging Face, MaxText, NVIDIA NeMo, and TensorRT-LLM.

Gemma models can run on various platforms, from laptops and workstations to Google Cloud, with optimisation for industry-leading performance on NVIDIA GPUs and Google Cloud TPUs.

This development comes after Google recently introduced Gemini 1.5 with a 1 million token context window — the largest ever seen in natural language processing models. In contrast, GPT-4 Turbo has a 128K context window, and Claude 2.1 has a 200K context window.

The post Google Challenges Meta’s Llama 2 with Lightweight Open Source LLM, Gemma appeared first on Analytics India Magazine.

Image by Author

When working with data, especially if using our beloved Python language, the dictionary stands out as a fundamental data structure, ready to uncover its data to those who know how to unlock it.

A dictionary in Python is a collection that is both unordered and mutable, designed to store data values like a map. Unlike other Data Types that hold only single values as elements, Dictionary holds pairs of keys and values separated by colons, a “:” element.

This Key-value pairs structure provides a way to store data so that it can be efficiently retrieved by key rather than position.

However, most times an unwanted KeyError when looking for a key can break our whole execution. This is why this guide attempts to shed some light and explain some effective ways to access dictionaries avoiding the break of our execution.

Understanding the Python Dictionary

Imagine a dictionary as a dynamic storage system, where each item you wish to store has a unique identifier or 'key' that leads you directly to it.

Image by Author

In Python, dictionaries are declared with curly brackets {}, with keys and their corresponding values separated by colons “:”, and each pair separated by commas.

Here's a simple representation:

# Creating a simple dictionary with keys and values  salaries = {     'Data Scientist': 100000,     'Data Analyst': 80000,     'Data Engineer': 120000}    print(salaries)

Creating a dictionary is just the beginning. The true utility of dictionaries is realized when retrieving and manipulating this stored data.

The Common Pitfall

A common approach to accessing a value in a dictionary is by using the key name within square brackets:

# Accessing a value using the key  print(salaries['Data Scientist'])  # Outputs: 100000    print(salaries['Professor'])  # This will raise a KeyError, as 'Professor' key doesn't exist

This method seems straightforward until you encounter a key that doesn't exist within the dictionary, leading to a KeyError.

This is a common issue that can complicate larger projects.

The Safer Approaches

To avoid KeyError, you might consider using if statements or try-except blocks to handle missing keys.

These methods, while functional, can become cumbersome with more complex code. Fortunately, Python offers more elegant solutions, mainly two:

• the get() method
• the setdefault() method

Embracing the get() Method

The get() method is a more efficient way to retrieve values from a dictionary. It requires the key you're searching for and allows an optional second parameter for a default value if the key is not found.

# Using get() to safely access a value  salary = salaries.get('Data Scientist', 'Key not found')  print(salary)  # Outputs: 30    # Using get() with a default value if the key doesn't exist  salary = salaries.get('Professor', 'Key not found')  print(salary)  # Outputs: 30

This is the most straightforward way to access a dictionary while ensuring there won’t be a KeyError. Having a default alternative is a safe way to make sure everything is in order.

However, this method does not alter our dictionary, and sometimes, we require the dictionary to store this new parameter.

This leads us to the second approach.

Leveraging the setdefault() Method

For scenarios where you not only want to retrieve a value safely but also update the dictionary with new keys, setdefault() becomes invaluable.

This method checks for the existence of a key and if absent, adds the key with the specified default value, effectively modifying the original dictionary.

# Using setdefault() to get a value and set it if not present  salary = salaries.setdefault('Professor', 70000)  print(salary)  # Outputs: 70000 since 'Professor' was not in the dictionary    # Examining the dictionary after using setdefault()  print(salaries) # Outputs: {'Data Scientist': 100000, 'Data Analyst': 80000, 'Data Engineer': 120000, 'Professor': 70000}

Examining playersHeight after using setdefault() will show the newly added keys with their default values, altering the original dictionary structure.

Final Recommendations

The choice between get() and setdefault() depends on your specific needs. Use get() when you simply need to retrieve data without altering the original dictionary. Opt for setdefault() when your task requires adding new entries to the dictionary.

Breaking old habits may require some effort, but the transition to using get() and setdefault() can significantly enhance the robustness and readability of your Python code.

As you integrate these methods into your programming practice, you'll quickly appreciate their efficiency and the seamless way they handle potential errors, making them indispensable tools in your Python arsenal.

The pivotal role of get() and setdefault() emerge as great ways to handle such formats and access our dictionaries.

I hope this guide was useful, and next time you are dealing with dictionaries, you can do it in more effectively.

You can go check the corresponding Jupyter Notebook in the following GitHub repo.

Josep Ferrer is an analytics engineer from Barcelona. He graduated in physics engineering and is currently working in the Data Science field applied to human mobility. He is a part-time content creator focused on data science and technology. You can contact him on LinkedIn, Twitter or Medium.

More On This Topic

• Do You Read Excel Files with Python? There is a 1000x Faster Way
• Predict Customer Churn (the right way) using PyCaret
• Converting JSONs to Pandas DataFrames: Parsing Them the Right Way
• 15 Trending MLOps Talks You can Access for Free at ODSC East 2022
• Hone Your Data Skills With Free Access to DataCamp
• Data access is severely lacking in most companies, and 71% believe…

Google has unveiled Gemma, a new family of open models, leveraging the research and technology behind the existing Gemini models. The Gemma open models are released in two sizes, Gemma 2B and Gemma 7B, each offering pre-trained and instruction-tuned variants.

Users can start working with Gemma today using free access in Kaggle and a free tier for Colab notebooks. Additionally, first-time Google Cloud users can avail themselves of $300 in credits. Researchers can also apply for Google Cloud credits of up to $500,000 to accelerate their projects.

Gemma outperforms Llama 2 on several benchmarks, including MMLU, HellaSwag, and HumanEval.

To support developer innovation and responsible use, Google is also providing a Responsible Generative AI Toolkit alongside the models. This toolkit includes essential tools for creating safer AI applications with Gemma, offering guidance and support for developers.

To facilitate widespread adoption, Gemma is compatible with major frameworks, including JAX, PyTorch, and TensorFlow through native Keras 3.0. The release includes ready-to-use Colab and Kaggle notebooks, integration with popular tools such as Hugging Face, MaxText, NVIDIA NeMo, and TensorRT-LLM.

Gemma models can run on various platforms, from laptops and workstations to Google Cloud, with optimisation for industry-leading performance on NVIDIA GPUs and Google Cloud TPUs.

This development comes after Google recently introduced Gemini 1.5 with a 1 million token context window — the largest ever seen in natural language processing models. In contrast, GPT-4 Turbo has a 128K context window, and Claude 2.1 has a 200K context window.

The post Google Challenges Meta’s Llama with Lightweight Open Source LLM, Gemma appeared first on Analytics India Magazine.

Google’s recent release, Gemini 1.5, with a 1M context length window, has sparked a new debate about whether RAG (Retrieval Augmented Generation) is still relevant or not. LLMs commonly struggle with hallucination. To address this challenge, two solutions were introduced, one involving an increased context window and the other utilising RAG.

Lately, several developers have been experimenting with Gemini 1.5. “I uploaded the Great Gatsby with two alterations (mentioning an ‘iPhone-in-a-box’ and a ‘laser lawnmower’). Gemini nails it (& finds one more thing). Claude does but hallucinates. RAG doesn’t work,” Ethan Mollick, a professor at Wharton, wrote on X.

Another X user, Mckay Wrigley, fed an entire biology textbook into Gemini 1.5 Pro, which consisted of 491,002 tokens. He asked it three extremely specific questions, and it provided each answer 100% correctly.

“Gemini 1.5 Pro is still underhyped. I uploaded an entire codebase directly from GitHub, and all of the issues, including Vercel AI SDK. Not only was it able to understand the codebase, but it also identified the most urgent issue and implemented a fix. This changes everything,” wrote Sully Omar, co-founder and chief at Cognosys.

The three examples above prove that Gemini 1.5, with its extensive context window, successfully retrieves crucial information within the document. However, this doesn’t portray the limitation of RAG.

Comparing Apples and Oranges

Many are still confused about the distinction between RAG and the context window. The context window limits the model to information within a given text span, while RAG extends the model’s capabilities to external sources, vastly widening the scope of accessible information.

Taking notice of the hype on the internet, Oriol Vinyals, VP of Research and Deep Learning team lead at Google DeepMind, voiced his opinion, saying, “RAG (retrieval-augmented generation) isn’t done for, even though we can handle 1M or more tokens in context now. In fact, RAG has some nice properties that can enhance (and be enhanced by) long context.”

“RAG allows you to find relevant information, but the way the model accesses it may be too restrictive due to compression. Long context may help bridge that gap, similar to how L1/L2 cache & main memory work together in modern CPUs,” he added.

A larger context window allows LLMs to consider more text and thus generates more accurate and coherent responses, particularly for complex and long sentences. However this doesn’t mean that the model won’t hallucinate.

According to a paper titled ‘Lost in the Middle: How Language Models Use Long Contexts,’ published by researchers from Stanford University, UC Berkeley, and Samaya AI, LLMs exhibit high information retrieval accuracy at the document’s start and end. However, this accuracy declines in the middle, especially with increased input processing.

RAG Survives the Day

“The worst take that I have seen these past few days is that long context models like Gemini 1.5 will replace RAG,” wrote Elvis Saravia, co-founder at DAIR.AI explaining that long-context LLMs work great with static information (books, video recordings, PDFs, etc.) but they are yet to be battle-tested on highly evolving information and knowledge.

He further added that to tackle these types of problems, one could potentially combine RAG and long-context LLMs to build a robust system that effectively and efficiently retrieves and performs large-scale analysis of key historical information.

“We will make progress towards addressing some of the challenges like “lost in the middle” and handling more complex structured and dynamic data but we still have a long way to go,” he said. Saravia added that different families of LLMs will help solve different types of problems. “We need to move on from this idea that there will be one LLM that will rule all.”

Without a doubt, Gemini 1.5 outperforms Claude 2.1 and GPT-4 Turbo as it can assimilate entire code bases, process over 100 papers, and various documents, but it surely hasn’t killed RAG.

The post Did Google Gemini 1.5 Really Kill RAG? appeared first on Analytics India Magazine.