PhD in Natural Language Processing | Applied Generative AI Researcher & Developer | Bridging AI Innovation with Real-World Solutions | Lead Data Scientist @GovTech
After the release of the RAG playbook, the AI CapDev team in GovTech Singapore’s AI Practice Group continues to explore ways to improve LLM apps.
One of the questions we are often asked is, “Beyond prompt engineering and few-shot learning, is there a way to improve the quality of LLMs’ responses in context-based question answering?”
We took the fine-tuning approach and trained two models, Llama-3-8B and Gemma-2-9B, to reply to parliamentary questions. How do the results compare to that of pre-trained instruction-tuned models?
Read my new blog post to find out!
https://lnkd.in/gWgqfxfW#aritficialintelligence#generativeai#llm#finetuning#evaluation
A great article on the end-to-end, nuts and bolts, process of building an AI solution. From the dirty data, to the cost of deployment.
Any moment now, an "LLMOps" charlatan will pop out of the weeds...
PhD in Natural Language Processing | Applied Generative AI Researcher & Developer | Bridging AI Innovation with Real-World Solutions | Lead Data Scientist @GovTech
After the release of the RAG playbook, the AI CapDev team in GovTech Singapore’s AI Practice Group continues to explore ways to improve LLM apps.
One of the questions we are often asked is, “Beyond prompt engineering and few-shot learning, is there a way to improve the quality of LLMs’ responses in context-based question answering?”
We took the fine-tuning approach and trained two models, Llama-3-8B and Gemma-2-9B, to reply to parliamentary questions. How do the results compare to that of pre-trained instruction-tuned models?
Read my new blog post to find out!
https://lnkd.in/gWgqfxfW#aritficialintelligence#generativeai#llm#finetuning#evaluation
PhD in Natural Language Processing | Applied Generative AI Researcher & Developer | Bridging AI Innovation with Real-World Solutions | Lead Data Scientist @GovTech
After the release of the RAG playbook, the AI CapDev team in GovTech Singapore’s AI Practice Group continues to explore ways to improve LLM apps.
One of the questions we are often asked is, “Beyond prompt engineering and few-shot learning, is there a way to improve the quality of LLMs’ responses in context-based question answering?”
We took the fine-tuning approach and trained two models, Llama-3-8B and Gemma-2-9B, to reply to parliamentary questions. How do the results compare to that of pre-trained instruction-tuned models?
Read my new blog post to find out!
https://lnkd.in/gWgqfxfW#aritficialintelligence#generativeai#llm#finetuning#evaluation
Legal Engineer and Legal AI expert at Wolters Kluwer Hungary Ltd.; Postgraduate student at Eötvös Loránd University
I am thrilled to announce the publication of our latest article together with MONTANA Knowledge Management and Wolters Kluwer Hungary in the Hungarian Journal of Legal Studies, titled "Increasing access to legal information with unsupervised solutions".
🔍 Why This Matters: In today's digital age, an enormous amount of legal data is generated daily, presenting significant challenges in processing and tracking changes, especially for those without a legal background. Understanding different legal information is crucial for ensuring access to justice.
🤖 Our Approach: We explored how a specific type of machine learning tool, unsupervised learning, can assist in navigating vast amounts of unstructured legal information. These sophisticated algorithms can identify patterns in documents and automatically categorize them, which significantly aids in understanding and accessibility.
💡 Key Findings: Our research demonstrated that unsupervised solutions efficiently group documents based on their similarities and differences. Additionally, they facilitate understanding each category by aggregating defining words, phrases, and word pairs, making the legal landscape more navigable for everyone.
🙌 Acknowledgments: Special thanks to Zsolt Ződi, Miklos Konczol, and the editorial team at HJLS for this incredible opportunity to contribute to the field.
You can find our open-access publication here: https://lnkd.in/dPYgzpAV#LegalTech#MachineLearning#AccessToJustice#UnsupervisedLearning
We are also working on applying LLMs as a judge for the relevance of information retrieval/search systems for data. So, it is great to see this type of application of LLMs gaining more attention!
Gathering relevance dataset to measure improvements or adverse impacts of changes to search/retrieval systems is challenging. LLMs can scale this for us, provided sufficient human baselines are established!
#AI#AIforData#llm4data
𝗨𝘀𝗶𝗻𝗴 𝗟𝗟𝗠-𝗮𝘀-𝗮-𝗷𝘂𝗱𝗴𝗲 🧑⚖️ 𝗳𝗼𝗿 𝗮𝗻 𝗮𝘂𝘁𝗼𝗺𝗮𝘁𝗲𝗱 𝗮𝗻𝗱 𝘃𝗲𝗿𝘀𝗮𝘁𝗶𝗹𝗲 𝗲𝘃𝗮𝗹𝘂𝗮𝘁𝗶𝗼𝗻
Evaluating LLM outputs is often hard, since many tasks require open-ended answers for which no deterministic metrics work: for instance, when asking a model to summarize a text, there could be hundreds of correct ways to do it. The most versatile way to grade these outputs is then human evaluation, but it is very time-consuming, thus costly.
🤔 Then 𝘄𝗵𝘆 𝗻𝗼𝘁 𝗮𝘀𝗸 𝗮𝗻𝗼𝘁𝗵𝗲𝗿 𝗟𝗟𝗠 𝘁𝗼 𝗱𝗼 𝘁𝗵𝗲 𝗲𝘃𝗮𝗹𝘂𝗮𝘁𝗶𝗼𝗻, by providing it relevant rating criteria? 👉 This is the idea behind LLM-as-a-judge.
⚙️ To implement a LLM judge correctly, you need a few tricks.
✅ So 𝗜'𝘃𝗲 𝗷𝘂𝘀𝘁 𝗽𝘂𝗯𝗹𝗶𝘀𝗵𝗲𝗱 𝗮 𝗻𝗲𝘄 𝗻𝗼𝘁𝗲𝗯𝗼𝗼𝗸 𝘀𝗵𝗼𝘄𝗶𝗻𝗴 𝗵𝗼𝘄 𝘁𝗼 𝗶𝗺𝗽𝗹𝗲𝗺𝗲𝗻𝘁 𝗶𝘁 𝗽𝗿𝗼𝗽𝗲𝗿𝗹𝘆 𝗶𝗻 𝗼𝘂𝗿 𝗛𝘂𝗴𝗴𝗶𝗻𝗴 𝗙𝗮𝗰𝗲 𝗖𝗼𝗼𝗸𝗯𝗼𝗼𝗸! (you can run it instantly in Google Colab)
➡️ 𝗟𝗟𝗠-𝗮𝘀-𝗮-𝗷𝘂𝗱𝗴𝗲 𝗰𝗼𝗼𝗸𝗯𝗼𝗼𝗸: https://lnkd.in/e7XEdhSi
The Hugging Face Cookbook is a great collection of notebooks demonstrating recipes (thus the "cookbook") for common LLM usages. I recommend you to go take a look!
➡️ 𝗔𝗹𝗹 𝗰𝗼𝗼𝗸𝗯𝗼𝗼𝗸𝘀: https://lnkd.in/eWu6aDny
Thank you Maria Khalusova for your support!
Great quick starter for using LLM as a judge!
Many works have explored using LLM as a judge for point-wise, pair-wise and ranking evaluations. Based on my recent reading, here are some limitations identified and methods proposed to address them:
✅ Position bias -> swapping positions
✅ Knowledge bias -> reference guided judge
✅ Limited reasoning capabilities -> COT + reference guided judge + Actor Critic
✅ Fine-tuning for evaluation further addresses the limitation and improved reliability. Examples include JudgeLM and Arena fine-tuned Vicuna.
🤔 Although promising results have been shown for LLM evaluations, challenges remain for multimodal LLM-as-a-judge, especially gaps from human preferences in point-wise scoring and ranking evaluations. The main issues identified include biases, hallucination and inconsistencies.
References:
https://lnkd.in/g6JBbzK3https://lnkd.in/gTgPRARXhttps://lnkd.in/gWAkSfjV
𝗨𝘀𝗶𝗻𝗴 𝗟𝗟𝗠-𝗮𝘀-𝗮-𝗷𝘂𝗱𝗴𝗲 🧑⚖️ 𝗳𝗼𝗿 𝗮𝗻 𝗮𝘂𝘁𝗼𝗺𝗮𝘁𝗲𝗱 𝗮𝗻𝗱 𝘃𝗲𝗿𝘀𝗮𝘁𝗶𝗹𝗲 𝗲𝘃𝗮𝗹𝘂𝗮𝘁𝗶𝗼𝗻
Evaluating LLM outputs is often hard, since many tasks require open-ended answers for which no deterministic metrics work: for instance, when asking a model to summarize a text, there could be hundreds of correct ways to do it. The most versatile way to grade these outputs is then human evaluation, but it is very time-consuming, thus costly.
🤔 Then 𝘄𝗵𝘆 𝗻𝗼𝘁 𝗮𝘀𝗸 𝗮𝗻𝗼𝘁𝗵𝗲𝗿 𝗟𝗟𝗠 𝘁𝗼 𝗱𝗼 𝘁𝗵𝗲 𝗲𝘃𝗮𝗹𝘂𝗮𝘁𝗶𝗼𝗻, by providing it relevant rating criteria? 👉 This is the idea behind LLM-as-a-judge.
⚙️ To implement a LLM judge correctly, you need a few tricks.
✅ So 𝗜'𝘃𝗲 𝗷𝘂𝘀𝘁 𝗽𝘂𝗯𝗹𝗶𝘀𝗵𝗲𝗱 𝗮 𝗻𝗲𝘄 𝗻𝗼𝘁𝗲𝗯𝗼𝗼𝗸 𝘀𝗵𝗼𝘄𝗶𝗻𝗴 𝗵𝗼𝘄 𝘁𝗼 𝗶𝗺𝗽𝗹𝗲𝗺𝗲𝗻𝘁 𝗶𝘁 𝗽𝗿𝗼𝗽𝗲𝗿𝗹𝘆 𝗶𝗻 𝗼𝘂𝗿 𝗛𝘂𝗴𝗴𝗶𝗻𝗴 𝗙𝗮𝗰𝗲 𝗖𝗼𝗼𝗸𝗯𝗼𝗼𝗸! (you can run it instantly in Google Colab)
➡️ 𝗟𝗟𝗠-𝗮𝘀-𝗮-𝗷𝘂𝗱𝗴𝗲 𝗰𝗼𝗼𝗸𝗯𝗼𝗼𝗸: https://lnkd.in/e7XEdhSi
The Hugging Face Cookbook is a great collection of notebooks demonstrating recipes (thus the "cookbook") for common LLM usages. I recommend you to go take a look!
➡️ 𝗔𝗹𝗹 𝗰𝗼𝗼𝗸𝗯𝗼𝗼𝗸𝘀: https://lnkd.in/eWu6aDny
Thank you Maria Khalusova for your support!
My comment on this article:
It seems to me that this approach resembles how we humans memorize facts. For example, if we are taking a history class, we might first read a chapter to get an overview. Then, knowing that the history test will requires us recalling exact dates, we might then go back over the chapter looking for dates which we would write on a piece of paper with what happen on that date. The night before the test, we would review the dates on our paper. We don’t run 100 epochs over the data, but enough times to get a good enough grade on the test.
But if the history professor told us that it would be an open book test, then we would bring in the book so that we could look up the exact dates. Then instead of memorizing dates, we would focus on understanding the concepts but also how to quickly find the dates in the book. This seems more like the RAG (Retrieval-Augmented Generation) approach being used by many LLMs where a specific database is accessed or by doing queries into the internet.
I suspect the ability to do both approaches will be found in future LLMs.
🌐 Exciting News for Tech Enthusiasts and Content Creators!
🚀 I'm thrilled to share my latest video: "Transform Your Ideas into Stunning AI-Powered PDF Reports | Easy Guide with LLM & Jinja2." This cutting-edge tutorial is more than just a guide; it's a glimpse into the future of automated document creation.
In this video, I delve into the fascinating world of AI Language Models, focusing on Mistral and MOE (Mixture of Experts), and demonstrate how to seamlessly merge these with Jinja2 for converting HTML documents into professional PDF reports. Whether you're a researcher, business professional, or a content creator, this is your opportunity to revolutionize how you approach document creation.
💡 What You'll Discover
The ins and outs of using AI for creating enriched content.
Easy-to-follow steps for converting HTML to PDF using Jinja2.
Practical tips to enhance the quality of your AI-generated reports.
This video is not just informative; it's a hands-on guide for anyone looking to harness the power of AI in their work. From generating research papers to crafting detailed business reports, the possibilities are endless.
🤝 Connect & Share
I'm excited to hear your thoughts and see how you apply these techniques. Let's start a conversation! Feel free to comment below, share your experiences, and let's explore the limitless potential of AI together.
🔔 Stay Updated
Follow my page for more content on AI innovations and tech tutorials. Together, we can stay at the forefront of technological advancements.
#AI#DocumentCreation#TechInnovation#Jinja2#LLM#MOE#MistralAI#ContentCreation#BusinessTechnology#FutureOfWork#LinkedInTechCommunity#CodeWithPrince
👍 Your Support Matters
If you find this video helpful, please like, share, and comment. Let's spread the word and empower more professionals with the knowledge to leverage AI in their fields!
https://lnkd.in/drDhuYNy
Final Year Computer Science Student at NTU
1moVery interesting! Did you consider using fine tuned LLM-as-a-Judge models like Prometheus 2.0 for the evaluation process?