Check out the winning solutions from our recent ML competition with LMSYS! The LMSYS Chatbot Arena has become an important platform for testing new LLMs. Using data from these evaluations, we worked with LMSYS to run a ML competition that challenged participants to predict which responses users would prefer between LLM chatbots. 👇Here are a few key takeaways we learned. - LLMs are effective at text-classification tasks (even in constrained environments) - Google ’s Gemma-2-9b was found to be especially capable at understanding user preference data - Reward/ranking models were found to be more useful than base models (e.g. models from Ai2's RewardBench). - Knowledge distillation was found to be an effective strategy to get optimal performance out of mid-sized models - Pseudo-labeling was found to be an effective strategy for generating additional training data in order to improve model performance - Averaging together multiple LoRAs was found to give more robust predictions as compared to working with only a single adapter Dig deeper by: 🔎Reviewing all of the solution write-ups on Kaggle https://lnkd.in/diN2TD_4 🎥Watching the winners calls from a couple of the top winners: https://lnkd.in/eSYmnx9q Congratulations and thank you to all the participants! 👏👏👏
About us
Kaggle provides cutting-edge data science, faster and better than most people ever thought possible. We have a proven track-record of solving real-world problems across a diverse array of industries including pharmaceuticals, financial services, energy, information technology, and retail. Kaggle offers both public and private data science competitions and on-demand consulting by an elite global talent pool.
- Website
-
http://www.kaggle.com
External link for Kaggle
- Industry
- IT Services and IT Consulting
- Company size
- 11-50 employees
- Headquarters
- San Francisco, California
- Type
- Privately Held
- Founded
- 2010
- Specialties
- open data, predictive modeling, machine learning, and data science
Locations
-
Primary
188 King Street #502
San Francisco, California 94107, US
Employees at Kaggle
Updates
-
🤖 Now on #KaggleModels! Qwen2 is a new series of Qwen LLMs optimized for language understanding & generation, multilingual capability, coding, mathematics, and reasoning. Available in 5 sizes ranging from 0.5 to 72 billion parameters. Learn more: https://lnkd.in/eJEir2kt
QwenLM | Qwen2 | Kaggle
kaggle.com
-
Kaggle reposted this
The youth mental health crisis has many causes, but one rising concern is technology. To address this, the Child Mind Institute and Kaggle have launched a competition using deidentified data from the Child Mind Institute Healthy Brain Network study to develop a predictive model that can identify early signs of problematic internet use (PIU) in young people using physical activity and fitness measures. The competition, sponsored by Dell Technologies and NVIDIA, launched on September 19th and will run for 3 months. This effort aims to create tools for clinicians to better understand and manage PIU. Learn more: https://lnkd.in/eCpwTC59 #MentalHealth #DataScience
Child Mind Institute and Kaggle Launch Competition to Predict Teen Internet Addiction - Child Mind Institute
childmind.org
-
Kaggle reposted this
✨ Our new Gemma 2 2B variant is fine-tuned for Japanese → https://goo.gle/4gO44Pv We're releasing this model along with training materials to serve as practical examples and learning resources, and to empower developer communities worldwide to adapt Gemma to their own languages. The new model runs on mobile phone hardware quickly, performing Japanese-language tasks comparable to GPT 3.5. You can run it on your mobile device using MediaPipe, and starting today, you can download Gemma 2’s model weights from Kaggle and Hugging Face. Let’s bridge communication gaps together. Join the “Unlocking Global Communication with Gemma” competition and fine-tune Gemma 2 to your language. Read the full blog for competition details.
-
Competition Launch Alert! Unlocking Global Communication with Gemma AI language translation is a strong example of where lightweight open models shine – there are over 7,000 languages, countless cultural nuances, and translations need to work quickly on smartphones. Today, Google released a new Gemma 2 2b variant fine-tuned for Japanese with training materials as a learning resource. Their goal is to empower communities worldwide to adapt Gemma to their own languages, using their deep understanding of their languages and cultures. To further support this effort, we've partnered with Google to launch the Unlocking Global Communication with Gemma competition. Participants are invited to fine-tune Gemma 2 for their languages and share their knowledge through reproducible notebooks that explore elements like language fluency, literary traditions, historical texts, and more. 💰 $150,000 Prize Pool ⏰ Entry Deadline: 1/14/25 Your work will help unlock the full potential of language AI and create a more connected and understanding world. Learn more at https://lnkd.in/eazkU7Ja
Google - Unlock Global Communication with Gemma
kaggle.com
-
Kaggle reposted this
Join the “Unlocking Global Communication with Gemma” competition on Kaggle → https://goo.gle/3Bs1w9J The competition was launched alongside the release of a Japanese variant of Gemma 2 2B, and training lessons for tuning Gemma to your own language, as part of #GemmaDeveloperDay in Tokyo, to inspire devs to build Gemma models for many languages. Start building and submit your entry. 🛠️
-
🤖 Now on #KaggleModels! AI at Meta latest foundation model, Llama 3.2, introduces small and medium-sized vision LLMs (11B and 90B), alongside lightweight, text-only models (1B and 3B) optimized for edge and mobile devices. Leading in openness, modifiability, and cost efficiency! Learn more: https://lnkd.in/eDXqQCaZ
Meta | Llama 3.2 | Kaggle
kaggle.com
-
📣There are 2 new types of achievements you can earn on Kaggle – Awards and Badges! Awards: Awards recognize all the awesome work Kagglers do outside of the progression system like winning analytics competitions, hosting competitions, speaking at special Kaggle events, etc. Badges: Badges are focussed on lightweight fun. You can earn Badges for trying out different features across Kaggle and for your activity across the site. We hope these new achievements better recognize all the amazing contributions the community makes on Kaggle and empower you to customize your Kaggle profile into an excellent portfolio. Full announcement: https://lnkd.in/e2J3yGgw
-
Kaggle reposted this
Data Science & Modelling Senior Expert at illimity Bank | MBA | Book Author @ Wiley, Packt, Manning | 3x Kaggle Grandmaster | GDE in AI and Machine Learning
Small Language Models (SLM), which are generally partially open-sourced (their weight and architecture are public), allow you to build your own applications locally, customizing models to your specific tasks without fear of any sudden breaking change that may make redundant and unusable all your work. In this project, I illustrate how you can effectively use a Gemma 2 2B model (one of the nimblest SLMs around) and have it learn new specialized skills. The idea is to train a model to assist you in reading and learning from new AI/ML papers. You can read this new article at: https://lnkd.in/dZWGb9fD You can also try the code and personalize it for your projects on this Kaggle notebook: https://lnkd.in/dfwz_RJ5 #AISprint #Google #Gemma #Kaggle #AI
-
🤖 Now on #KaggleModels! Simple yet fast, SAM 2 is a state-of-the-art model for segmenting objects in images and videos. Highly performant across a wide range of tasks and visual domains. Learn more: https://lnkd.in/eRKvT4Qc
Meta | Segment Anything 2 | Kaggle
kaggle.com