Add other projects

Author: BoyeGuillaume

docs/projects.md

@@ -58,6 +58,22 @@
 
     [:octicons-arrow-right-24: See Distillation of Medical LLMs information](#distillation-of-medical-llms)
 
+- :material-head-dots-horizontal:{ .lg .middle } __Meditron reasoning__
+
+    ---
+
+    Integrate reasoning through unsupervised reinforcement learning into Meditron aiming to further elevate its performance and decision-making abilities.
+
+    [:octicons-arrow-right-24: See Meditron reasoning information](#meditron-reasoning)
+
+- :material-language-html5:{ .lg .middle } __MOOVE__
+
+    ---
+
+    *MOOVE* (Massive Open Online Validation and Evaluation) is a large-scale, participatory evaluation platform designed to collect, structure, and analyze expert feedback on the outputs of clinical large language models (LLMs).
+
+    [:octicons-arrow-right-24: See MOOVE information](#moove-massive-open-online-validation-and-evaluation)
+
 </div>
 
 ## MMORE & Mirage
@@ -141,102 +157,42 @@ __Required Experience:__
 - Solid Python programming
 - Familiarity with training and evaluating neural networks
 - Basic understanding of language models and knowledge distillation techniques
-<!-- ## 1. Meditron
-
-- **MultiMeditron**
-
-  This project is about making Meditron multimodal: the user can provide Meditron with medical images, in addition to text. Work is two-fold: adapting the codebase of Meditron to make it have a multimodal architecture, and making the "expert" models that process the images and make embeddings fed to Meditron.
-
-  Contact: Michael Zhang (michael.zhang@epfl.ch)
-
-- **Fine-tuning multimodal models for the medical use**
-
-  This project aims to fine-tune generalist SOTA multimodal models (Qwen3 Omni, Llava, Llama4,...) with our medical multimodal data mixture. The goal is to build the best open-weights medical multimodal model according to the standard benchmark
-
-  Contact: Michael Zhang (michael.zhang@epfl.ch)
-
-- **Meditron Reasoning**
-
-  This project aims to improve our training pipeline by integrating novel reinforcement learning approaches, notably using GRPO algorithms. This is the continuation of a previous project conducted in this area, and we plan to expand the existing work to enhance our project performances (add MultiMeditron for multi-modal reasoning).
-
-  Contact: Guillaume Boyé (guillaume.boye@epfl.ch)
-
-- **Polyglot Meditron & Giving Meditron a Voice**
-
-  Speaking English is nice, most content online is in English. Having a performant LLM for medical tasks formulated in English is useful. But not enough! In low-resource settings and even in most places of the globe, people usually prefer using their first language rather than English.
-
-  There are many people around the world who even though cannot read, seek healthcare information and guidance. Currently, medical LLMs, even those that are multi-modal, are usually constrained to a few languages thereby limiting their application in this particular use-case of healthcare question answering. The main objective of this project is to extend the multi-lingual speech capabilities of our Meditron model to ensure that it is more accessible to people around the world.
-
-  This project aims at making Meditron models more proficient in other languages, with a focus on low-resource languages. In written and spoken speech. Work is needed, since having a polyglot base model is generally not enough: popular models do not have a focus on low-resource languages, and there is also a need to make sure to teach the model non-English medical terminology.
-
-  Contact: Fabrice Nemo (fabrice.nemo@epfl.ch) & David Sasu (david.sasu@epfl.ch)
-
-- **NeuroMeditron**
-
-  NeuroMeditron develops robust multimodal models for dementia prediction using voice and typing dynamics from the mPower dataset. The project focuses on handling missing modalities through advanced fusion strategies, enabling reliable patient-level monitoring. A proof-of-concept “Neuro Expert” adapter will integrate these digital biomarkers into MultiMeditron.
-
-  Contact: Arianna Francesconi (arianna.francesconi@epfl.ch)
-
-  
-- **Meditron-4: Clinical feedback alignment and SOTA dev**
-
-  Meditron-4 is the next iteration of Meditron, designed to close the gap between medical knowledge and guideline-faithful, clinically contextualized behavior. While Meditron-3 is now lagging behind state-of-the-art, Meditron-4 will deliver an open-source fine-tuning and evaluation pipeline and the best clinically aligned model we can produce on top of leading open medical and general base models—while also pushing small, offline-capable models (e.g., MedGemma 4B, LFM-2) for low-resource deployment.
-
-  Contact: Xavier Theimer-Lienhard (xavier.theimer-lienhard@epfl.ch)
-
-## 2. MMORE
-
-  MMORE stands for Massive Multimodal Open RAG & Extraction, it is our Python library for a scalable multimodal pipeline for processing, indexing, and querying multimodal documents.
-
-  [GitHub repo](https://github.com/swiss-ai/mmore)
-
-  Contact: Fabrice Nemo (fabrice.nemo@epfl.ch)
-
-## 3. Moove
-
-  The [moove](https://jointhemoove.org) is a collaborative platform where experts and communities co-design and validate AI models. The initiative focuses on aligning large language models with real-world standards, ensuring they are transparent, safe, and context-aware. It is already partnered with institutions such as CHUV, ICRC, the Gates Foundation and many hospitals around the world.
-
-  If you want to help us make the moove even greater, don't hesitate to join!
-
-  Note that the project is software-engineering focused.
-
-Contact: Bryan Gotti (bryan.gotti@epfl.ch)
-
-## 4. HIC-Lab AI Bootcamp
-
-Very cool project about teaching the basics of AI applied to healthcare. The target audience is healthcare workers and computer scientists in Rwanda. Our work in LiGHT is to improve the content of the bootcamp so that students learn better, and mentor students there, guide them throughout their completion of the bootcamp.
-
-Contact: Fabrice Nemo (fabrice.nemo@epfl.ch)
-
-## 5. CHIT-CHAT
-1. Embedding Humanitarian Principles in LLM Development
-
-LLMs are usually not deployed for humanitarian applications since they are not intentionally designed to align to humanitarian values. This project therefore aims to develop a framework / checklist for LLM development and evaluation that can be applied in the creation and testing of Humanitarian-focused LLMs.
-
-Contact: David Sasu (david.sasu@epfl.ch)
-
-## 6. PRISM-AI
 
-PRISM-AI leverages the PRISM dataset on pregnancy reference intervals to benchmark traditional ML/DL models against Large Language Models for risk prediction in maternal health. The project explores fine-tuning strategies and novel optimization methods (e.g., DPO/GRPO) to assess whether LLMs can provide clinically meaningful improvements over established approaches.
+## Meditron reasoning
 
-Contact: Arianna Francesconi (arianna.francesconi@epfl.ch)
+*This project is supervised by Guillaume Boyé and Lars Klein*
 
-## 7. Multimodal Learning from Voice and Keyboard Dynamics for Early Alzheimer’s Diagnosis
+Reasoning has been a significant breakthrough in advancing the capabilities of
+large language models in recent years. It has consistently demonstrated its ability
+to enhance decision-making processes within these systems. The objective of this project
+is to integrate reasoning through unsupervised reinforcement learning into Meditron
+aiming to further elevate its performance and decision-making abilities.
 
-This project develops deep learning model to detect early Alzheimer’s disease from typing and voice signals. Students will design a multimodal models (RNNs for typing and CNN/ViT for voice) to capture motor and speech patterns linked to cognitive decline, comparing modality contributions and model interpretability.
+__Completed:__
 
-Contact: Arianna Francesconi (arianna.francesconi@epfl.ch)
+- Integrated VERL on the cluster with distributed training on multi-node with appropriate docker image
+- Docker image for SGLang inference
+- LLM-as-a-judge based reward
+- Distributed setup
+- Prototype dataset and prototype reward function
+- Prototype support for multiturn and tooling for python execution
 
-## 8. Cross-Disease Voice Prognosis: Parkinson and ALS Audio Modeling
+__Possible Tasks:__
 
-Voice changes are early markers of neurodegenerative diseases. This project trains deep learning models on Parkinson’s voice recordings (mPower) and tests cross-disease generalization on ALS speech data, exploring transfer learning and shared vocal biomarkers across disorders.
+- Experiment with new datasets and reward modeling for reasoning tasks to enhance model generation
+- Explore additional RL algorithms and architecture for improving capabilities (multi-agent setup)
+- Expand the tool based and introduce RAG system to improve the observability of the reasoning
+- Benchmark model performance on complex tasks
 
-Contact: Arianna Francesconi (arianna.francesconi@epfl.ch)
+__(Required) Experience:__
 
-## 9. Balancing Time-Series Health Data Across Diseases
+- Strong knowledge of __Python__, __PyTorch__ experience is a plus
+- Experience on distributed infrastructure using __SLURM__, working with server is a plus
+- __Linux__ knowledge (for building Docker image, GLHF)
+- Knowledge of reward modeling is a plus
 
-This project extends the [IMBALMED method](https://www.sciencedirect.com/science/article/pii/S0895611125000382) for class balancing in time-series models (LSTM/GRU) and benchmarks it against standard techniques such as SMOTE or focal loss. Students will analyze cross-disease robustness and ensemble diversity, building a reproducible benchmark for temporal health data.
+## MOOVE: Massive Open Online Validation and Evaluation
 
-Contact: Arianna Francesconi (arianna.francesconi@epfl.ch)
+*This project is supervised by Fay Elhassan and Karian For*
 
- -->
+*MOOVE* (Massive Open Online Validation and Evaluation) is a large-scale, participatory evaluation platform designed to collect, structure, and analyze expert feedback on the outputs of clinical large language models (LLMs). Built in collaboration with clinicians and healthcare institutions across diverse geographies including Sub-Saharan Africa, South Asia, Latin America, and Europe MOOVE is the first multilingual, context-sensitive evaluation environment tailored to healthcare AI systems in low- and middle-income as well as high-resource settings.