Kili Technology’s Post

🎬 Webinar Replay: Quality Metrics explained in less than 4min 👀 ✅ Now that you have selected your framework to perform RLHF on your LLM, you need to measure quality metrics to keep track of the fine-tune. 📉📈 A few best practices to adopt: 🔸 QA Score 🔸 Metrics Trackers (Behavioural distributions, Bias, and Data diversity) 🔸 Agreement (Preference ranking) 😉 Everything is explained by Paul G. in this video 👇 🔥 Need to deep-dive the subject a bit more? Our full webinar recap is available here: https://lnkd.in/efbTz67Q #webinar #AI #RLHF

You ask, we answer. 🙂↕️ 🤔 ‘’Why use RLHF, why don’t we ask LLM itself to tell which response is better?’’ 👉 While using an LLM to evaluate responses (often called "AI feedback" or "LLM as a judge") can be useful, especially for bootstrapping or when human feedback is unavailable, RLHF with human feedback is generally preferred because: ✨ Humans can provide more nuanced and contextually appropriate judgments, especially for complex or specialized tasks. ✨ Human feedback helps avoid potential biases or limitations present in evaluating LLM. ✨ RLHF allows for the optimization of specific metrics or business goals that an LLM might not inherently understand. We hope this clarifies your concerns! 😌 We finished a webinar recap that summarises everything we talked about. Have a look at it 👇 #AI #LLM #RLHF

Kili Technology is proud to be part of the elite group of companies belonging to the NVIDIA Inception Program 🥳🔥 The NVIDIA Inception group has been formed to encourage start-ups to drive change across industries worldwide. We thank NVIDIA for their trust, and we’re excited to keep building the best datasets to fuel the world’s most powerful models. We’re so proud to see how well Kili Technology has been doing recently. And this is only the beginning 💪 Stay up to date on LinkedIn! ✅ #news #AI #program

🚀💭 In our latest webinar, the Adaptive ML team dove deep into the advantages of preference tuning via Reinforcement Learning from Human Feedback (RLHF) over traditional supervised fine-tuning. ✅ Here’s what Andrew Jardine shared about it: - Superior Results: RLHF can significantly enhance performance on specific tasks compared to fine-tuning alone. - Efficient Data Collection: Gathering preference data is often easier and more impactful than creating perfect training outputs. - Continuous Improvement: Models can be refined continuously through ongoing user feedback, enabling better adaptation. - Cost-Effective: Smaller models fine-tuned with RLHF can outperform larger models, potentially saving computing costs and latency. - Flexibility: RLHF allows optimization beyond mimicking outputs, targeting user satisfaction or business goals. 🔥 The webinar highlighted all the advantages of reinforcement learning. 👀 Check out the full scope through our webinar recap article with Paul G., Andrew Jardine and Daniel Hesslow. (link in the comments) 👇 #webinar #replay #RLHF

🔥 Huge congratulations to our CFO, Mounia MIR, for being recognized as one of the 2024 “CFOs Under 40 at the forefront” by Décideurs Corporate Finance! 🎉 At Kili Technology, we’re proud to have modern thinkers like Mounia on our team, driving positive change and inspiring innovation. 🌟 We believe in encouraging bright minds to step forward and lead the way! Way to go, Mounia! 👏💪 Check out her interview with the link in the comment below 👇 #Leadership #CFO #Innovation #TeamKili

🎬 Webinar Replay: Quality People and Quality Workflow 👀 💪 Building a robust RLHF dataset requires both skilled people and efficient workflows, combined with the right tools and feedback mechanisms to ensure continuous quality improvements. But what are the two parameters about and how important are they? 🤔 Everything you need to know with Paul G. in our Webinar Replay Part 2: #RLHF #replay #LLM

Unfolding Meta's LLaMA 3.1: What does the Future Hold? 🔍 Meta's LLaMA 3.1 is here, showcasing an unprecedented 405 billion parameters. But what does this mean for the future of AI and large language models? 🤔 In our blog, we delve into: - ⚙️ The technical prowess of LLaMA 3.1: Understanding the architecture that powers this massive model. - 💻 Data strategies: How Meta’s use of diverse datasets elevates model performance and accuracy. - 🤖 Impact on the AI ecosystem: Potential applications and the broader implications for AI development. ✅ Whether you're an AI enthusiast or a professional in the tech space, this article offers a comprehensive overview of where AI is headed. Don’t miss out on these insights 👇 #AI #Llama31 #TrainingData

In an industry where bigger often seems better, SmolLM a small language model by Hugging Face focuses on efficiency without compromising performance. 😮💨 🔥 After rigorous testing, it has outperformed its peer models in key areas: - 🔍 Common Sense Reasoning: SmolLM excels in tasks requiring everyday knowledge, demonstrating its practical AI capabilities. - 🌍 World Knowledge: Its curated datasets allow SmolLM to answer questions and make inferences across a wide range of topics with impressive accuracy. - 💻 Technical Proficiency: SmolLM shines in Python programming, generating precise code snippets that are invaluable for developers and educators. SmolLM shows that focused training and curated data can achieve exceptional results. Read our full article to know more! 😉👇 #AI #Innovation #SmolLM

🏃♂️ Dynamic Labeling vs Static labeling 🧍♀️ 👀 In reinforcement learning, a key aspect to success is having enough of the right data to improve your model. In some cases, you may not have the right quantity, hence, a dynamic labeling workflow is needed. What is dynamic labeling you may ask? 🤔 🎬 Our previous webinar with Paul G. and Adaptive ML tackles this workflow. Check out his explanation below: 👇 #webinar #replay #labeling

Meta's LLaMA 3.1: A New Frontier in LLM’s! 🚀 Meta's latest innovation, the LLaMA 3.1, is breaking barriers in AI with its staggering 405 billion parameters. This isn't just a leap in scale—it's a evolution in how we approach natural language processing. 🤖 In our latest blog, we explore it’s: - Model Architecture: How Meta’s approach differs from competitors. - Training Data: The role of diverse, high-quality data in enhancing performance. - Applications: Real-world use cases and what LLaMA 3.1 means for industries across the board. 🌐 If you're involved in AI, tech, or just curious about the future of language models, this is a must-read. 😉 Discover what makes LLaMA 3.1 a game-changer 👇 #AI #TrainingData #LLMs