Vincent Nguyen on LinkedIn: Embarrassingly small English-to-German Model (258M parameters) with… (original) (raw)
Vincent Nguyen’s Post
Seedfall | ex-president Ubiqus | Machine learning
7mo Edited
Co-training an NMT Model along a Comet-like Estimator Exciting news! Alongside François Hernandez, we are gearing up to unveil the successor of the OpenNMT-py toolkit. Our collaborative efforts have led to a significant overhaul of various features, culminating in a groundbreaking achievement. We successfully trained a compact NMT model with just 258M parameters, yielding state-of-the-art results by co-training it with a comet-like Estimator. Our methodology leverages established techniques, reminiscent of distillation, as detailed in this research paper: https://lnkd.in/evatHuUH. Furthermore, our approach bears resemblance to a recent publication: https://lnkd.in/eUPPtmNg. Notably, this innovative process is underpinned by the Comet Architecture developed by Unbabel, accessible here: https://lnkd.in/eqm2h8YE. Stay tuned as we unravel the intricacies of our journey and share insights on how we achieved this remarkable feat.#NMT #MachineLearning #OpenNMT #CometEstimator #Innovation
Also, NeuralDesktop by Panagiotis Kanavos which provides a 100% offline solution with cutting edge features will integrate this innovation. Make sure you visit his website.
Special thanks to Ricardo Rei who gave me some insights on Comet.
What are the next languages? This sounds exciting
👏Congrats! This could be the much-needed game-changer. 🤞 First AI post in months that made me click +Follow. 🔭
To view or add a comment, sign in
``
More Relevant Posts
-
Language Industry Researcher | Founder @ Custom.MT
7mo
The Magician of Machine Translation Returns! Vincent Nguyen worked for 22 years to build Ubiqus into France's top-3 largest translation company. In the later years, Ubiqus' AI department and their GPU servers with open cases and constant coil whine resided on top of a skyscraper in la Defense, Paris's extravagantly expensive financial quarter. Symbolically, the linguists and project managers continued working from the ground and the 1st floors. Vincent sold his mighty company in 2022 and secluded himself in an ivory tower. Having the millions, the talent, and the organizational capability to join the GenAI race, he sidestepped the hype of the blooming Parisian scene and continued to work on machine translation in secret. Today, Vincent emerges from his secluded laboratory bearing a gift of magi - a small but powerful MT model that allegedly may rival DeepL in quality. It's open alongside the methodology, and other companies may now productize it, for example, for on-device machine translation. I'm sure Apple will be curious. In an age where the world is crazy about large and generic language models, is there a place for conventional and small MT? What's next for mysterious Vincent? If you were as strong, where would you focus your energies and talents?
Seedfall | ex-president Ubiqus | Machine learning
7mo Edited
Co-training an NMT Model along a Comet-like Estimator Exciting news! Alongside François Hernandez, we are gearing up to unveil the successor of the OpenNMT-py toolkit. Our collaborative efforts have led to a significant overhaul of various features, culminating in a groundbreaking achievement. We successfully trained a compact NMT model with just 258M parameters, yielding state-of-the-art results by co-training it with a comet-like Estimator. Our methodology leverages established techniques, reminiscent of distillation, as detailed in this research paper: https://lnkd.in/evatHuUH. Furthermore, our approach bears resemblance to a recent publication: https://lnkd.in/eUPPtmNg. Notably, this innovative process is underpinned by the Comet Architecture developed by Unbabel, accessible here: https://lnkd.in/eqm2h8YE. Stay tuned as we unravel the intricacies of our journey and share insights on how we achieved this remarkable feat.#NMT #MachineLearning #OpenNMT #CometEstimator #Innovation
Embarrassingly small English-to-German Model (258M parameters) with State-of-the-art results link.medium.com
To view or add a comment, sign in -
CSE||SIH'23 Finalist|| Competitive Programmer || Codeforces (Max- 1145) || 2 ⭐ (Codechef) || C++ || Data Science || Machine Learning Developer
1w
Day 66 of GFG Problem Solving Challenge: Equilibrium Point Store the the sum in a prefix and suffix array and for each i check if prefix[i-1]==suffix[i+1] if the condition satisfies return i. Time Complexity: O(n) Space Complexity: O(n).#gfg #gfg160 #geekstreak2024 #womenintech
To view or add a comment, sign in -
Training & Support Director at KOREC Group
5mo
Trimble Perspective really enables an efficient laser scanning workflow, with in-field registration and georeferencing of laser scan data. We've added a series of training videos to our Perspective playlist. Even if you know Perspective well, it may be worth taking a look just to make sure you aren't missing a trick. #trustkorec KOREC Group https://lnkd.in/ePp3kKns
Perspective - Creating Annotations
https://www.youtube.com/
To view or add a comment, sign in
-
CSE||SIH'23 Finalist|| Competitive Programmer || Codeforces (Max- 1145) || 2 ⭐ (Codechef) || C++ || Data Science || Machine Learning Developer
1w
Day 60 of GFG Problem Solving Challenge: Count the number of possible triangles Sort the vector. Iterate j from n-1 to j=2 and for each j apply two pointers at left=0 and right=j-1. If ((arr[left]+arr[right])<=arr[j]) then do left++ , else do count+=(right-left) and right--. By this, we can count all possible triangles that can be formed. Time Complexity: O(n^2) Space Complexity: O(1).#gfg #gfg160 #geekstreak2024 #womenintech
To view or add a comment, sign in - #ModelingToolkit now supports sampled-data systems! Discover how it helps #simulate complex systems with mixed continuous-discrete processes, perfect for control systems like #autopilots and #signal #processing. https://lnkd.in/evEbQ4xw #JuliaLang #simulation #juliacon #conference
Modeling and simulation of sampled-data systems | Bagge Carlson | JuliaCon 2024
https://www.youtube.com/
To view or add a comment, sign in
-
"Aeronautical Engineer Turned Data Scientist | Applying Analytical Precision to Data & Technology"
2mo
Euclideian Algorithm deals with the with principle between two number gcd and divides their difference allows to reduce the iterations and untill the number becomes zero Case 2 Series 2
To view or add a comment, sign in -
"Aeronautical Engineer Turned Data Scientist | Applying Analytical Precision to Data & Technology"
2mo
Euclideian Algorithm deals with the with principle between two number gcd and divides their difference allows to reduce the iterations and untill the number becomes zero Case 1 Series 1
To view or add a comment, sign in -
Regional Manager at Emlid | MBA | Geospatial Solutions | Novice Business Strategist
4mo
Emlid Flow’s 𝗕𝗔𝗦𝗘 𝗦𝗛𝗜𝗙𝗧 feature lets you measure points when your benchmark is out of reach. Check out this impressive video by Mangoesmapping, presented by Alistair Hart. (And yes, the feature is free)https://lnkd.in/gcRWBpEi
Using Base Shift in Emlid Flow
https://www.youtube.com/
To view or add a comment, sign in