Prabakaran Chandran’s Post

"Has the horse already bolted on data sovereignty and ownership for te reo Māori?" We had some thought-provoking kōrero last night at our Kōrero AI Technical Panel Discussion. 🤖 Ngā mihi aroha ki a koutou ki ngā korokoro taniwha; 🤖 Jahminique Grace - Māori Data Analyst and AI Specialist 🤖 Punahamoa Walker - Accomplished AI Researcher 🤖 Michael Puhara - Self-taught Developer and AI Enthusiast Some of the kōrero shared from our speakers and the audience; 💚 We need to be at the forefront of creating our own models, involving iwi, hapū and those communities who the model is built for, in the process. 💚 Some barriers to creating an LLM could include; data quality, open vs closed source, being without access to technology or people in rural communities and larger corporates not having the kaupapa at the centre of what they do. 💚 Might we be able to use AI as a tool for language preservation? Only if we work with the impacted communities to understand their needs in terms of preservation, rather them tell them that we are here to preserve their language. 💚 Is anyone thinking about the intersection of AI and linguistics? There are so many dialects (within every language, not just te reo Māori) that need to be accounted for when building these models. Effortless code-switching between languages was also talked about. An "AI linguistic notation" was mentioned. Engari, I'll leave that kōrero for those who studied linguistics, nē. 💚 The best way to get into AI? Just start! Subscribe to YouTube channels, watch TikTok educational videos, curate your LinkedIn feed and just try it out. Although the horse may (or may not) have bolted, there is still time to reign it in so-to-speak, with an indigenous saddle me kī. 🐴 Once again, a special thank you to everyone who took time out of their evening to dail-in, ask pātai and share whakaaro. Mīharo ake nei! See you at the next one. 💻 🐉 💚 #TechTaniwha #AI #ArtificialIntelligence #Aotearoa

Read our top ACL publication on AI safety! We are a group of researchers who have a record of publishing in top AI places. If you have a challenging AI problem related to AI governance (evaluation/safety/alignment), we can solve it for you! Email: admin@walled.ai Website: walled.ai

ASMv2: Revolutionizing Relation Comprehension in Visual AI The All-Seeing Project V2 introduces a new model, ASMv2, designed for advanced image understanding, particularly in identifying and comprehending object relations. It combines text generation, object localization, and relational understanding into a single task, aimed at enhancing Multi-modal Large Language Models (MLLMs) capabilities. The project includes a specialized ReC dataset for training and a novel CRPE benchmark for evaluation, showcasing significant improvements in relation understanding, evidenced by its outperformance of existing models. This work aims to push the boundaries of AI towards greater interpretative and relational insight in visual data. #ASMv2 #AllSeeingProject #VisualAI #RelationComprehension #ObjectRelations #MultiModalLargeLanguageModels #MLLMs #ReCDataSet #CRPEBenchmark #AIInnovation #VisualData #Research #Interpretation

Here’s a little-known trick to boost your AI game: The language you use with Large Language Models (LLMs) directly impacts the quality of responses. And guess what? It’s not the model’s fault it's all about the data. Here’s the inside scoop LLMs are trained on a massive amount of English data, giving them a deeper understanding of context, tone, and accuracy. However, the quality can dip when you switch to languages with less training data (like Portuguese). Even if you're working in a different language, try this method to get better responses: ✅ Translate Your Prompt into English Start with English. It’s the goldmine for rich, nuanced answers. ✅ Generate the Response Let the LLM do its magic in English, where it’s most effective. ✅ Translate Back to Your Target Language Take that high-quality response and shift it back to your desired language: Voilà, a more detailed and accurate output. Why it Works More data means more precision. By using English, you're leveraging the LLM’s strongest dataset. This leads to better answers, even in specialized or complex topics. When to Use This ↳ Specialized industries ↳ Tackling complex queries ↳ Writing for multilingual audiences The next time you work with an LLM in a multilingual context, consider translating your prompt to English first. It’s a simple switch, but the results can be game-changing. #AI #generativeAI #LLMs #multilingual #promptengineering #translationHacks #automation

𝐇𝐮𝐠𝐠𝐢𝐧𝐠 𝐅𝐚𝐜𝐞'𝐬 𝐌𝐮𝐥𝐭𝐢𝐥𝐢𝐧𝐠𝐮𝐚𝐥 𝐒𝐩𝐞𝐞𝐜𝐡-𝐭𝐨-𝐒𝐩𝐞𝐞𝐜𝐡 𝐩𝐢𝐩𝐞𝐥𝐢𝐧𝐞 𝐢𝐬 𝐦𝐚𝐤𝐢𝐧𝐠 𝐚 𝐬𝐩𝐥𝐚𝐬𝐡! 🌊 This revolutionary tool allows you to effortlessly switch languages mid-conversation, thanks to its lightning-fast 100ms delay. Imagine having GPT4o-like experiences on your device, no matter where you are or who you're talking to. 🌎 Building on the incredible success of our GitHub repository (2600 stars and counting!), they have expanded the library to support multiple languages. 𝐖𝐚𝐧𝐭 𝐭𝐨 𝐭𝐫𝐲 𝐢𝐭 𝐨𝐮𝐭? Specify a language: Add the flag --language fr for French, or any other supported language. Let the system decide: Skip the flag, and let the pipeline automatically detect the language. The future of multilingual communication is here! 🌍 #AI #MachineLearning #SpeechRecognition #Multilingual #HuggingFace #GPT4o #LanguageBarrier

Quantization is an indispensable technique for serving Large Language Models (LLMs) and has recently found its way into LoRA fine-tuning. However, in such cases, it is common to observe a consistent gap in the performance on downstream tasks between full fine-tuning and quantization plus the LoRA fine-tuning approach. 𝐋𝐨𝐟𝐭𝐐: 𝐋𝐨𝐑𝐀-𝐅𝐢𝐧𝐞-𝐓𝐮𝐧𝐢𝐧𝐠-𝐀𝐰𝐚𝐫𝐞 𝐐𝐮𝐚𝐧𝐭𝐢𝐳𝐚𝐭𝐢𝐨𝐧 𝐟𝐨𝐫 𝐋𝐚𝐫𝐠𝐞 𝐋𝐚𝐧𝐠𝐮𝐚𝐠𝐞 𝐌𝐨𝐝𝐞𝐥𝐬 a new paper published last week focused on the scenario where quantization and LoRA fine-tuning are applied together on a pre-trained model. LoftQ (LoRA-Fine-Tuning-aware Quantization) is a novel quantization framework that simultaneously quantizes an LLM and finds a proper low-rank initialization for LoRA fine-tuning. Such an initialization alleviates the discrepancy between the quantized and full-precision model and significantly improves the generalization in downstream tasks. This method is evaluated on natural language understanding, question answering, summarization, and natural language generation tasks and it outperforms existing quantization methods, especially in the challenging 2-bit and 2/4-bit mixed precision regimes. ➡ Paper Link: https://lnkd.in/dSPcbxrb ⭐ Subscribe to To Data & Beyond to receive similar content: https://lnkd.in/dVkD-7WT

Diving into the world of Language Model (LLM) Functions! 🧑💻🌐 👇 As #LLMs become more and more powerful and get more and more use cases, I thought it's important to have a structured approach for levareging their capabilities. This is where #LLM Functions come in. An LLM Function is essentially a task-specific wrapper around a language model and it includes: 1. An LLM instance (the actual #AI model) 🤖 2. A prompt template (instructions for the specific task) 📝 Encapsulating these components into one re-usable LLM Function you can get a clear interface for performing different LLM-specific task, such as: - Text Summarization 📄➡️📃  - Question Answering ❓➡️💬 - Language Translation 🌐➡️🌍 On a high level, the power of LLM Functions lies in their modularity and reusability, but you're welcome to my Medium post to dive deeper into LLM Functions and better understand how it actually works: https://lnkd.in/gVvNxg8B

Got the recent work accepted at ACL 2024-main track. When I started my PhD about four years ago, I felt a lot happier about top AI publications, regardless of their impact beyond the research community. Now it feels like I need to make them valuable to the general public. So, I will aim to work in directions that have a wider impact, more specifically in the AI safety domain. RESTA is one such work, receiving scores of 4, 4, and 4 (/5) from the reviewing committee, placing it in the top 3% of papers submitted to the ARR. Paper: Language Models are Homer Simpson! Safety Re-Alignment of Fine-tuned Language Models through Task Arithmetic Impact: Prevent safety compromise of LLMs due to enterprise-specific fine-tuning. Link: https://lnkd.in/gtbRY_5k Code: https://lnkd.in/geAtnctr Hugginface (safety evaluation multilingual dataset): https://lnkd.in/gEzMV3Pp More about the approach?-https://lnkd.in/gb_y62hX Grateful to my collaborators Duc Anh Do and Soujanya Poria. AI safety needs a lot of exploration-including discussions on its definition, evaluation, and alignment. #AISafety #NLProc #ACL2024

Embarking on the transformative journey of language generation with cutting-edge LLMs (𝐋𝐚𝐫𝐠𝐞 𝐋𝐚𝐧𝐠𝐮𝐚𝐠𝐞 𝐌𝐨𝐝𝐞𝐥𝐬) Exciting news! Presenting our latest exploration, "Mastering the Art of Language Generation: A Guide to 𝐋𝐚𝐫𝐠𝐞 𝐋𝐚𝐧𝐠𝐮𝐚𝐠𝐞 𝐌𝐨𝐝𝐞𝐥𝐬 Parameters." This guide is a treasure trove of insights that have revolutionized my digital communication and creativity. Delving deep into this guide, I've uncovered the profound impact of four crucial parameters shaping the linguistic prowess of 𝐋𝐚𝐫𝐠𝐞 𝐋𝐚𝐧𝐠𝐮𝐚𝐠𝐞 𝐌𝐨𝐝𝐞𝐥𝐬: 🔑 Temperature: A game-changer, adjusting the temperature unlocks creativity and innovation, revealing remarkable word choices that elevate surprise in our creations. 🔑 Top_P: Broadening the adventure into word selection, a higher Top_P value introduces delightful randomness, enriching our textual landscapes with diverse vocabulary. 🔑 Frequency Penalty: Effectively addressing repetition, elevating the frequency penalty ensures vibrant and captivating language, avoiding the mundane echo of redundancy. 🔑 Presence Penalty: Sharpening focus and coherence through the presence penalty ensures content relevance. Fine-tuning this parameter aligns our words closely with intended topics. Exploring these parameters, especially through examples starting with 'A,' has revealed the vast capabilities and versatility of 𝐋𝐚𝐫𝐠𝐞 𝐋𝐚𝐧𝐠𝐮𝐚𝐠𝐞 𝐌𝐨𝐝𝐞𝐥𝐬. It's a clear demonstration of harnessing AI's power to enrich our linguistic expressions. I'm thrilled to invite you to delve into this guide and embark on your language generation exploration with 𝐋𝐚𝐫𝐠𝐞 𝐋𝐚𝐧𝐠𝐮𝐚𝐠𝐞 𝐌𝐨𝐝𝐞𝐥𝐬. Let's leverage AI to elevate our creativity and communication! 🌍 Download this complete Guide for Free by Joining our vibrant AI community. For a deep dive into language generation mechanics and to start your journey, check the link in the comments below. Together, let's explore, create, and inspire! Follow Decoding Data Science for insightful posts #ai #tech #llm #generativeai

🌟 𝐇𝐨𝐥𝐢𝐬𝐭𝐢𝐜 𝐄𝐯𝐚𝐥𝐮𝐚𝐭𝐢𝐨𝐧 𝐨𝐟 𝐋𝐚𝐧𝐠𝐮𝐚𝐠𝐞 𝐌𝐨𝐝𝐞𝐥𝐬: 𝐀 𝐌𝐮𝐬𝐭-𝐑𝐞𝐚𝐝! 🌟 I've recently come across a fantastic paper titled "𝐇𝐨𝐥𝐢𝐬𝐭𝐢𝐜 𝐄𝐯𝐚𝐥𝐮𝐚𝐭𝐢𝐨𝐧 𝐨𝐟 𝐋𝐚𝐧𝐠𝐮𝐚𝐠𝐞 𝐌𝐨𝐝𝐞𝐥𝐬" by Stanford researchers, and I can't help but share it with everyone. This paper is incredibly comprehensive, providing deep insights into evaluating the outputs of large language models (LLMs) across various scenarios. 🔍 𝐖𝐡𝐲 𝐘𝐨𝐮 𝐒𝐡𝐨𝐮𝐥𝐝 𝐑𝐞𝐚𝐝 𝐈𝐭: 𝐂𝐨𝐦𝐩𝐫𝐞𝐡𝐞𝐧𝐬𝐢𝐯𝐞 𝐂𝐨𝐯𝐞𝐫𝐚𝐠𝐞: The paper covers a wide range of evaluation metrics and methodologies, making it a valuable resource for understanding how to assess LLMs effectively. 𝐈𝐧𝐬𝐩𝐢𝐫𝐢𝐧𝐠 𝐁𝐮𝐬𝐢𝐧𝐞𝐬𝐬 𝐕𝐚𝐥𝐮𝐞: The insights from this paper can inspire new ways to generate business value by leveraging LLMs in innovative applications. 𝐆𝐫𝐞𝐚𝐭 𝐒𝐭𝐚𝐫𝐭𝐢𝐧𝐠 𝐏𝐨𝐢𝐧𝐭: Whether you're new to the field or an experienced professional, this paper serves as an excellent starting point for learning about the evaluation of LLM outputs. If you're interested in the future of AI and language models, this paper is a must-read! 📚✨ Read the paper here: 𝐇𝐨𝐥𝐢𝐬𝐭𝐢𝐜 𝐄𝐯𝐚𝐥𝐮𝐚𝐭𝐢𝐨𝐧 𝐨𝐟 𝐋𝐚𝐧𝐠𝐮𝐚𝐠𝐞 𝐌𝐨𝐝𝐞𝐥𝐬 (https://lnkd.in/e8G_wWQU) #AI #MachineLearning #LanguageModels #Evaluation #Research #Innovation