Daya Philip

Sarvam-M: A 24B Parameter Multilingual AI Model is finetuned Mistral-Small with Indic language desi nuances covering 11 major languages from india 🇮🇳 Key Features: • Dual-mode interface: Quick "non-think" responses + detailed "think" mode for complex reasoning 🧠 • Strong performance on math (GSM-8K) and coding (SWE-Bench) benchmarks 📊 • Supports both native Indic scripts and romanized text ✍️ • Built for real-world multilingual conversational agents 💬 I think it's a good start , instead of training model from scratch Sarvam focused on building indic dataset and model like in video shows good translingual conversational capability . Try it out via the Colab link - https://lnkd.in/gzcW4hru Playground - https://lnkd.in/gXijZrkY

27

1 Comment

Sarvam AI is outperforming large global LLMs on India-specific use cases. Across real-world India-focused tasks, Sarvam AI (fine-tuned for India) is delivering stronger performance than large general-purpose models — proving that targeted fine-tuning beats sheer parameter size. 🏆 Where Sarvam AI Outperformed: ✅ Vernacular Intent Detection (Hindi + regional mix) More accurate understanding of Hinglish, code-mixed queries, and regional phrasing. ✅ Regulatory & Compliance Q&A (India domain) Higher precision in interpreting India-specific policies and terminology. ✅ Localized Customer Support Responses More culturally aligned tone with fewer generic outputs. ✅ Cultural Context & Idiomatic Understanding Better handling of local expressions and socio-cultural nuances. Why This Is Happening Most global LLMs like those from OpenAI and Google (Gemini) are trained predominantly on large-scale English and Western-centric datasets. India’s linguistic landscape is fundamentally different: • 22+ official languages • Heavy code-mixing (Hinglish, Tanglish, etc.) • Deep cultural and regulatory context • Region-specific communication styles This gap is one reason why adoption of generic global LLMs in India can be slower for core business workflows. India needs models built for India’s data, languages, and users. Task-specific fine-tuning on curated India-centric datasets enables: • Higher accuracy • Lower latency • Better cost efficiency • Stronger real-world alignment 🧠 Key Insight: For country-specific and domain-heavy applications, alignment > size. Kudos to Sarvam for leading India’s LLM movement in the right direction. #AI #IndiaAI #LLM #MachineLearning #FineTuning #GenAI #DigitalIndia

20

Why does spectral clustering outperform K-means? How do eigenvectors reveal complex cluster structures? Rukshan Pramoditha presents a thorough explainer on a powerful machine learning approach.

18

Does your data make traditional clustering approaches like K-means less effective? Rukshan Pramoditha proposes an alternative: spectral clustering. Read his explainer to understand how (and why) it works.

12

🚀 Accelerating Responsible AI in India: IIT Bombay Releases 16 Open Datasets on AIKOSH 🇮🇳 . Indian Institute of Technology, Bombay Bombay has taken a significant step toward advancing responsible and inclusive AI by contributing 16 diverse datasets to AIKOSH, the Government of India’s official AI repository. These open datasets reflect India’s cultural and linguistic richness and are aimed at fueling research and innovation in Artificial Intelligence (AI) and Machine Learning (ML) — tailored for Indian use cases. 🔍 What’s Inside: Handwritten & printed Indian scripts Tables extracted from scanned documents Multi-language audio samples Drone surveillance imagery Visual question-answer datasets 🎯 Why It Matters: These datasets help AI models: Recognize and interpret Indian languages and handwriting Process documents and media with regional nuances Analyze audio-visual data in Indian contexts This initiative marks a crucial step toward building AI that works for India, by India — empowering researchers, developers, and institutions nationwide. 🔗 Explore the datasets on AIKOSH: https://lnkd.in/gJv4XS4j #AI #MachineLearning #ResponsibleAI #OpenData #IITBombay #AIKOSH #IndiaAI #Innovation #Research #DigitalIndia

31

5 Comments

📝 Announcing our EMNLP 2025 paper that introduces the Alignment Quality Index (AQI), a metric leveraging latent geometric representations and clustering indices to diagnose hidden misalignments in LLMs. 🔹 “𝐀𝐥𝐢𝐠𝐧𝐦𝐞𝐧𝐭 𝐐𝐮𝐚𝐥𝐢𝐭𝐲 𝐈𝐧𝐝𝐞𝐱 (𝐀𝐐𝐈) – 𝐁𝐞𝐲𝐨𝐧𝐝 𝐑𝐞𝐟𝐮𝐬𝐚𝐥𝐬: 𝐀𝐐𝐈 𝐚𝐬 𝐚𝐧 𝐈𝐧𝐭𝐫𝐢𝐧𝐬𝐢𝐜 𝐀𝐥𝐢𝐠𝐧𝐦𝐞𝐧𝐭 𝐃𝐢𝐚𝐠𝐧𝐨𝐬𝐭𝐢𝐜 𝐯𝐢𝐚 𝐋𝐚𝐭𝐞𝐧𝐭 𝐆𝐞𝐨𝐦𝐞𝐭𝐫𝐲, 𝐂𝐥𝐮𝐬𝐭𝐞𝐫 𝐃𝐢𝐯𝐞𝐫𝐠𝐞𝐧𝐜𝐞, 𝐚𝐧𝐝 𝐋𝐚𝐲𝐞𝐫 𝐰𝐢𝐬𝐞 𝐏𝐨𝐨𝐥𝐞𝐝 𝐑𝐞𝐩𝐫𝐞𝐬𝐞𝐧𝐭𝐚𝐭𝐢𝐨𝐧𝐬” 🔹 In collaboration with Manipal University Jaipur, LinkedIn, Indian Institute of Technology, Kharagpur, Indian Institute of Information Technology Guwahati, Texas A&M University, VIT Chennai, Evalueserve, New York University, Birla Institute of Technology and Science, Pilani (Pragya Lab) 🔹 Published in EMNLP 2025 Main 🔹 Paper: https://lnkd.in/g54bEBCu ✍🏼 Authors: Abhilekh Borah, Chhavi Sharma, Danush Khanna, Utkarsh Bhatt, GURPREET SINGH, Hasnat Md. Abdullah, Raghav Kaushik R, Vinija Jain, Jyoti Patel, Shubham Singh, Vasu Sharma, Arpita V., Rahul Raja, Aman Chadha, Dr. Amitava Das ➡️ 𝐊𝐞𝐲 𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬 𝐨𝐟 𝐨𝐮𝐫 𝐀𝐥𝐢𝐠𝐧𝐦𝐞𝐧𝐭 𝐐𝐮𝐚𝐥𝐢𝐭𝐲 𝐈𝐧𝐝𝐞𝐱 (𝐀𝐐𝐈): 🧬 𝑰𝒏𝒕𝒓𝒊𝒏𝒔𝒊𝒄 𝑳𝒂𝒕𝒆𝒏𝒕 𝑮𝒆𝒐𝒎𝒆𝒕𝒓𝒚 𝑴𝒆𝒕𝒓𝒊𝒄: AQI measures alignment by assessing how distinctly safe vs. unsafe prompts cluster in latent activations, using combination of Davies‑Bouldin, Dunn, Xie‑Beni, and Calinski‑Harabasz indices—making it invariant to decoding style and resilient to alignment faking. 🧠 𝑳𝒂𝒚𝒆𝒓𝒘𝒊𝒔𝒆 𝑷𝒐𝒐𝒍𝒆𝒅 𝑹𝒆𝒑𝒓𝒆𝒔𝒆𝒏𝒕𝒂𝒕𝒊𝒐𝒏 𝑳𝒆𝒂𝒓𝒏𝒊𝒏𝒈: we use a sparse learned pooling mechanism across transformer layers to capture alignment‑relevant abstractions without modifying the base model—enabling internal safety diagnostics. 🚨 𝑬𝒎𝒑𝒊𝒓𝒊𝒄𝒂𝒍 𝑭𝒂𝒊𝒍𝒖𝒓𝒆𝒔 𝒐𝒇 𝑩𝒆𝒉𝒂𝒗𝒊𝒐𝒓𝒂𝒍 𝑴𝒆𝒕𝒓𝒊𝒄𝒔: AQI uncovers misalignments that behavioral metrics like G‑Eval or refusal rates miss—especially under jailbreaks, safety‑agnostic fine‑tuning, and stochastic decoding, proving its value as an early‑warning alignment audit tool.

117

2 Comments

In this blog, I build on my previous work, where I used image texture attributes to augment image datasets (https://lnkd.in/gScutnbC). Here, I explore how to select the most effective feature extraction methods for image classification. While raw pixel data can be high-dimensional and noisy, feature extraction transforms this data into a more manageable and informative representation. Feature extraction involves identifying key characteristics within an image and converting them into numerical values. These characteristics can include texture, shapes, structures, edges, corners, histograms, and more. The Python package pfeats (https://lnkd.in/gBkiwyig) offers various feature extraction methods, but many more methods are also available. The challenge lies not just in applying feature extraction but in selecting the right features that provide the most informative insight. Simply put, we want features that can clearly distinguish one image from another, enabling machine learning models to perform at their best. To demonstrate the significance of selecting the most suitable features, I used the CIFAR-10 dataset, which comprises 10 classes of labeled images. I randomly selected 10 images from each class and computed 36 different features for each image. Then, I calculated the median value of each feature within each class. The figure at the top compares the median values across classes. While it is possible to identify differences, it remains challenging to definitively determine which features are most effective in distinguishing one labeled image from another. This is where more advanced visualization techniques become essential. I used the violin plot (shown in the figure at the bottom) as it illustrates the spread of the values for each extracted feature. The plot is organized and color-coded based on the interquartile range (IQR) as follows: - Green features have the largest IQR. These features vary significantly, making them excellent candidates for distinguishing between classes. - Yellow features have a moderate IQR. These features are also good options for classification. - Orange features have a low IQR. These features become less valuable as more features approach those colored in red. - Red features have the lowest IQR. These features are too similar across classes, making them the least useful for classification. This visualization procedure provides a clear, data-driven way to choose the best features for machine learning models, leading to better classification results. #machinelearning #imageprocessing #imageclassification #violinplot #visualanalytics #dataanalytics #python #datadiversity

12

1 Comment

Use Neo4j’s query language Cypher to manage and retrieve data stored in knowledge graphs. Write knowledge graph queries that find and format text data to provide more relevant context to LLMs for Retrieval Augmented Generation. Build a question-answering system using Neo4j and LangChain to chat with a knowledge graph of structured text documents.

4

Edge AI is gaining serious momentum 🚀 — and for good reason. Deploying models locally for data processing represents a meaningful shift in AI architecture — especially in regions where connectivity is inconsistent and data privacy concerns are rising, including across large parts of India. By moving inference closer to users, Edge AI can: Reduce latency and enable real-time performance ⚡ Strengthen data privacy by minimizing cloud dependency 🔒 Support functionality in low- or no-connectivity environments 📡 Lower cloud infrastructure costs for AI-driven products 💰 This shift can significantly expand access to AI-powered software for populations that were previously underserved 🌍. For startups, this is a strategic opportunity. With cloud costs being a long-standing pain point, optimizing for on-device deployment allows teams to focus more on product innovation and sustainable scale. While challenges remain — model optimization, hardware constraints, and deployment complexity — these are engineering problems that can be solved with the right architecture. Edge AI isn’t just a trend; it’s a structural shift in how intelligent systems will be built and deployed. Curious to hear your thoughts — is Edge AI becoming the default architecture for next-generation AI products?

25

1 Comment

🔧 How RLHF Is Built (Step-by-Step) 1️⃣ Pretrained Model Start with a large language model trained on massive data. Good at patterns — not yet aligned with human preference. 2️⃣ Supervised Fine-Tuning (SFT) Humans create example responses the model should learn to imitate. 3️⃣ Human Preference Data For each prompt, generate several responses. Humans rank them from best to worst. 4️⃣ Train a Reward Model This model learns to predict which answer humans prefer. 5️⃣ Reinforcement Learning (e.g., PPO) The main model is optimized to maximize the reward score while staying close to the original knowledge. The result is behavior that aligns with human expectations. 🌍 Real-Life Company Example: OpenAI’s ChatGPT -OpenAI uses RLHF to train ChatGPT so that its responses are more helpful, safe, and aligned with human preferences — not just statistically likely text. RLHF is the reason ChatGPT feels conversational, context-aware, and user-friendly rather than purely predictive. This isn’t just theory — it’s what makes modern assistants like ChatGPT useful in products across customer support, education, and enterprise workflows. ❤️ Why RLHF Matters Without RLHF: -Models are smart but can be tone-deaf. -They may produce harmful or unhelpful responses. With RLHF: -Responses feel natural and safe. -Behavior reflects human expectations. -User experience improves dramatically. RLHF doesn’t make models smarter — it makes them aligned. And in real-world AI products, alignment wins. #AI #MachineLearning #LLM #RLHF #Product #AIAlignment

4

Beyond Words: Using DETR Transformers to Decode Sign Language? Recently, I dove into the world of sign language detection models and came across a fascinating DETR transformer-based approach — a refreshing shift from the traditional OpenCV and TensorFlow pipelines. 📌 Github: https://lnkd.in/dys8uhBM 🎥 Demo: https://lnkd.in/dNz8bxKN What's exciting is the potential here — imagine extending this with a CTC-like loss approach (inspired by ASR) to generate live transcripts from detected signs. Coupled with WebRTC and TTS, we could be looking at real-time sign-to-speech translation platform. It’s just a thought for now, but with models like these becoming more accessible and powerful, bridging communication gaps may not be that far off. Loving this journey of continuous learning — from models to ideas. #MachineLearning #ComputerVision #SignLanguageRecognition #Transformers #AI #WebRTC #ASR #TTS #DeepLearning #OpenSource #LearningByDoing #SelfGrowth #ML #DL #WebRTC

15

🎉 Today’s AI Learning Update I delved into the renowned research paper “Attention Is All You Need,” a pivotal work that introduced Transformers 🔥. Here are the key insights I've gathered: 🔹 Pre-Transformers era saw challenges with slow sequence models like RNNs & CNNs due to their sequential or chunk-based text processing methods. 🔹 The breakthrough concept proposed discarding recurrence and convolution in favor of universal attention. 🔹 Attention facilitates direct connections between every word in a sentence, enabling seamless interactions like "The book … was amazing.” 🔹 The inception of the Transformer architecture featured: - An Encoder–Decoder framework. - Self-attention for word interactions. - Multi-head attention for diverse interpretations. - Feed-forward layers & residual connections for stability. 💡 Significance: Transformers offer: - Accelerated training through parallelization. - Enhanced handling of distant dependencies. - Pioneering translation outcomes, later underpinning GPT, BERT, LLaMA, and contemporary AI advancements. ✨ Personal Insight: This paper underscores the power of simplification, highlighting how groundbreaking ideas emerge by streamlining complexity (e.g., RNNs, CNNs) to focus on essentials. #AI #MachineLearning #DeepLearning #Transformers #AttentionIsAllYouNeed #LearningJourney

17

𝑬𝒙𝒄𝒊𝒕𝒆𝒅 𝒕𝒐 𝒔𝒉𝒂𝒓𝒆 𝒕𝒉𝒂𝒕 𝒐𝒖𝒓 𝒓𝒆𝒔𝒆𝒂𝒓𝒄𝒉 𝒑𝒂𝒑𝒆𝒓 𝒉𝒂𝒔 𝒃𝒆𝒆𝒏 𝒑𝒖𝒃𝒍𝒊𝒔𝒉𝒆𝒅 𝒃𝒚 𝑺𝒑𝒓𝒊𝒏𝒈𝒆𝒓 𝑵𝒂𝒕𝒖𝒓𝒆 𝒊𝒏 𝒕𝒉𝒆 𝒃𝒐𝒐𝒌 “𝑺𝒑𝒆𝒆𝒄𝒉 𝒂𝒏𝒅 𝑳𝒂𝒏𝒈𝒖𝒂𝒈𝒆 𝑻𝒆𝒄𝒉𝒏𝒐𝒍𝒐𝒈𝒊𝒆𝒔 𝒇𝒐𝒓 𝑳𝒐𝒘-𝑹𝒆𝒔𝒐𝒖𝒓𝒄𝒆 𝑳𝒂𝒏𝒈𝒖𝒂𝒈𝒆𝒔.” Our paper, “𝐎𝐩𝐭𝐢𝐦𝐢𝐳𝐢𝐧𝐠 𝐎𝐂𝐑 𝐌𝐨𝐝𝐞𝐥 𝐏𝐞𝐫𝐟𝐨𝐫𝐦𝐚𝐧𝐜𝐞: 𝐀 𝐂𝐨𝐦𝐩𝐚𝐫𝐚𝐭𝐢𝐯𝐞 𝐒𝐭𝐮𝐝𝐲 𝐨𝐟 𝐁𝐚𝐜𝐤𝐛𝐨𝐧𝐞 𝐀𝐫𝐜𝐡𝐢𝐭𝐞𝐜𝐭𝐮𝐫𝐞𝐬 𝐚𝐧𝐝 𝐇𝐲𝐩𝐞𝐫𝐩𝐚𝐫𝐚𝐦𝐞𝐭𝐞𝐫 𝐓𝐮𝐧𝐢𝐧𝐠” appears as a chapter in this Springer volume under section "Multimodality and Language Grounding to Vision and Robotics", part of the proceedings of the Third International Conference, SPELLL 2024, Chennai, India. This work focuses on enhancing OCR model accuracy and efficiency through comparative deep learning architectures and hyperparameter optimization contributing toward advancements in multimodal AI and language grounding for vision and robotics. Grateful to my co-authors , Dhanushkumar R, Harisudhan S, Dr. KARTHI M, and Kalpana B N for their collaboration and insights throughout this journey. 📘 Published in: Springer Nature 🔗 https://lnkd.in/gaeQU7eX 📄 Pages: 310–326 📚 ISBN (eBook): 978-3-032-05855-3 #Springer #SPELLL2024 #OCR #DeepLearning #ComputerVision #MultimodalAI #Research #AI #Publication #SpringerNature #MachineLearning #VisionandLanguage #LanguageTechnology

51

4 Comments

In this 20-minute talk, PolyMage Labs founder Uday Reddy Bondhugula talks about why and how one should build effective compilers for AI programming frameworks: https://lnkd.in/gWUQNUsF

79

📌 Epoch.AI just raised the bar: FrontierMath OpenProblems Instead of stopping at FrontierMath Tier 5, Epoch.AI went all in and released FrontierMath OpenProblems — a benchmark where models are asked to solve open research problems in mathematics. These are not contest puzzles. These are problems that professional mathematicians have worked on — and failed to solve so far. To make evaluation feasible, the benchmark only includes problems where a proposed solution can be automatically verified (which heavily limits the pool). For now, there are 14 problems, but submissions are open, so the set may grow — hopefully mathematicians will contribute. What kind of math? The problems span multiple areas, but the pilot set leans toward: combinatorics number theory These fields contain more statements that can be checked automatically. Human difficulty estimates Contributing mathematicians were asked: how many researchers have worked on the problem how long it might take humans to solve Some are estimated at 3–10 years of effort 😅 Current state of AI So far, today’s frontier models solve none of the tasks. Good luck to OpenAI, Google, and everyone else — if a breakthrough happens here, it will be a real milestone. #AI #Math #FrontierMath #EpochAI #LLM #Research #Benchmark #AutomatedReasoning #NumberTheory #Combinatorics #AGI

1