Add files using upload-large-folder tool

.gitattributes

@@ -33,3 +33,27 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+MediX-R1-8B-Q2_K.gguf filter=lfs diff=lfs merge=lfs -text
+MediX-R1-8B-Q3_K_S.gguf filter=lfs diff=lfs merge=lfs -text
+MediX-R1-8B-Q3_K_M.gguf filter=lfs diff=lfs merge=lfs -text
+MediX-R1-8B-Q3_K_L.gguf filter=lfs diff=lfs merge=lfs -text
+MediX-R1-8B-Q4_0.gguf filter=lfs diff=lfs merge=lfs -text
+MediX-R1-8B-Q4_1.gguf filter=lfs diff=lfs merge=lfs -text
+MediX-R1-8B-Q4_K_S.gguf filter=lfs diff=lfs merge=lfs -text
+MediX-R1-8B-Q4_K_M.gguf filter=lfs diff=lfs merge=lfs -text
+MediX-R1-8B-Q5_0.gguf filter=lfs diff=lfs merge=lfs -text
+MediX-R1-8B-Q5_1.gguf filter=lfs diff=lfs merge=lfs -text
+MediX-R1-8B-Q5_K_S.gguf filter=lfs diff=lfs merge=lfs -text
+MediX-R1-8B-Q5_K_M.gguf filter=lfs diff=lfs merge=lfs -text
+MediX-R1-8B-Q6_K.gguf filter=lfs diff=lfs merge=lfs -text
+MediX-R1-8B-Q8_0.gguf filter=lfs diff=lfs merge=lfs -text
+MediX-R1-8B-IQ3_S.gguf filter=lfs diff=lfs merge=lfs -text
+MediX-R1-8B-IQ3_M.gguf filter=lfs diff=lfs merge=lfs -text
+MediX-R1-8B-IQ4_XS.gguf filter=lfs diff=lfs merge=lfs -text
+MediX-R1-8B-IQ4_NL.gguf filter=lfs diff=lfs merge=lfs -text
+mmproj-MediX-R1-8b-F16.gguf filter=lfs diff=lfs merge=lfs -text
+assets/medix-r1_arch.png filter=lfs diff=lfs merge=lfs -text
+assets/reward_design_graph.png filter=lfs diff=lfs merge=lfs -text
+assets/microscopy_qualitative.png filter=lfs diff=lfs merge=lfs -text
+assets/xray_qualitative.png filter=lfs diff=lfs merge=lfs -text
+MediX-R1-8B-F16.gguf filter=lfs diff=lfs merge=lfs -text

MediX-R1-8B-F16.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:71bfba6d155e3f7a9eb90ea04900389b6b274da464c59143b46769bcaa2cb6c5
+size 16388045088

MediX-R1-8B-IQ3_M.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7157ae1d306b9a18d9bc24279d85645c9c671668176efeb6b4dfe1f8080a9113
+size 3896621088

MediX-R1-8B-IQ3_S.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:756eae053820440015fcbae6ca01c21d3edd95e8f90db02d41177918f00dc2c8
+size 3789666336

MediX-R1-8B-IQ4_NL.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a13c9fa58b68b7999a1f042a872878505337f724e5bfb64abb23aee21ad01569
+size 4818790432

MediX-R1-8B-IQ4_XS.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:66aff9e4de2b69f13c6afec950c035a10232aaa7759386a81c055768179f2afe
+size 4593297440

MediX-R1-8B-Q2_K.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e0e811b7a4b0315e6eeeeb81bb413c204ea143e934058157854e8313fe03e1b2
+size 3281733664

MediX-R1-8B-Q3_K_L.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6e1ba52e4aaa6e58a628e1c60e8c69d91a9defbc68c030316b54df75cead6393
+size 4431394848

MediX-R1-8B-Q3_K_M.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:256937adf103049202d2d5c86d90c74dfcdbfc307c536db61573677c401aa74c
+size 4124162080

MediX-R1-8B-Q3_K_S.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:666c38e6ffdbf29dc0b290696b87d8455487420fc6895cd662e27b19f288b18b
+size 3769612320

MediX-R1-8B-Q4_0.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ab882436f70e8365178e817b1dc6a79426137ae14f0370bc214c240080065670
+size 4774750240

MediX-R1-8B-Q4_1.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cfc54373889d37ac15b5fa3ffe7bc1681dc0db804cc7a737eead7ec66bbc5720
+size 5247756320

MediX-R1-8B-Q4_K_M.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c6728943aa60f090e6396159e5064d3290d03c05060e3129b9271f474f0d7e88
+size 5027784736

MediX-R1-8B-Q4_K_S.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5df976fe40738153d00173f1998e68369bba25c2c0f2754c48028c722bfe1641
+size 4802013216

MediX-R1-8B-Q5_0.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9d8930e32fdc54b1f42125cd68a8f932c0beae81b3d7942651a049b3544a7d6c
+size 5720762400

MediX-R1-8B-Q5_1.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:62b5b20bb7436d64028f17eed91c0d7e072a73aaef988644033f6d38fb0254ce
+size 6193768480

MediX-R1-8B-Q5_K_M.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8bf6849c38f39fdd341954aab94c2f415dea063df4ab56ac65c2193afa2a6afc
+size 5851113504

MediX-R1-8B-Q5_K_S.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c8c8bf5edfa53c981a99d6374204cea396a14980453a8d2acd11449ee3392a9a
+size 5720762400

MediX-R1-8B-Q6_K.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b4096c7bfa2ce29876be5765e4552a4de16d87afaa93ff41f9a7e5103f1b4241
+size 6725900320

MediX-R1-8B-Q8_0.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2245342474d96a124e07d6590492a3f946712745331a9174bcde526d2c094485
+size 8709519392

README.md

@@ -0,0 +1,159 @@
+---
+license: cc-by-nc-sa-4.0
+language:
+  - en
+tags:
+  - medical
+  - reinforcement-learning
+  - multimodal
+  - vision-language
+  - qwen3-vl
+pipeline_tag: image-text-to-text
+library_name: transformers
+---
+# MediX-R1: Open-Ended Medical Reinforcement Learning
+
+<p align="center">
+    <img src="assets/logo_white_no_bg.png" alt="MediX-R1" width="200">
+</p>
+
+<p align="center">
+    <img src="https://i.imgur.com/waxVImv.png" alt="MediX-R1">
+</p>
+
+#### [Sahal Shaji Mullappilly](https://scholar.google.com/citations?user=LJWxVpUAAAAJ&hl=en)\*, [Mohammed Irfan K](https://scholar.google.com/citations?user=GJp0keYAAAAJ&hl=en)\*, [Omair Mohamed](https://scholar.google.com), [Mohamed Zidan](https://scholar.google.com), [Fahad Khan](https://sites.google.com/view/fahadkhans/home), [Salman Khan](https://salman-h-khan.github.io/), [Rao Muhammad Anwer](https://scholar.google.com/citations?hl=en&authuser=1&user=_KlvMVoAAAAJ), and [Hisham Cholakkal](https://scholar.google.com/citations?hl=en&user=bZ3YBRcAAAAJ)
+
+\**Equally contributing first authors*
+
+#### **Mohamed Bin Zayed University of Artificial Intelligence (MBZUAI), UAE**
+
+[![Website](https://img.shields.io/badge/Project-Website-87CEEB)](https://medix.cvmbzuai.com)
+[![Paper](https://img.shields.io/badge/arXiv-Paper-red.svg)](https://arxiv.org/pdf/2602.23363)
+[![HuggingFace](https://img.shields.io/badge/HuggingFace-Page-F9D371)](https://huggingface.co/collections/MBZUAI/medix-r1)
+[![Leaderboard](https://img.shields.io/badge/MediX-Leaderboard-green)](https://medix.cvmbzuai.com/leaderboard)
+
+---
+
+## Overview
+
+MediX-R1 is an open-ended Reinforcement Learning (RL) framework for medical multimodal large language models (MLLMs) that enables clinically grounded, free-form answers beyond multiple-choice formats. MediX-R1 fine-tunes vision-language backbones with Group-Based RL and a composite reward tailored for medical reasoning: an LLM-based accuracy reward, a medical embedding-based semantic reward, and lightweight format and modality rewards that enforce interpretable reasoning.
+
+Despite using only ~50K instruction examples, MediX-R1 achieves excellent results across standard medical LLM and VLM benchmarks, outperforming strong open-source baselines.
+
+**Highlights:**
+- Our **8B** model achieves an overall average of **68.8%**, outperforming the much larger 27B MedGemma (68.4%).
+- Our **30B** model achieves the best overall score of **73.6%**, demonstrating the effectiveness of our composite reward design.
+
+---
+
+## Contributions
+
+- We introduce an **open-ended RL framework** for medical MLLMs that produces clinically grounded, free-form answers beyond MCQ formats.
+- We design a **composite reward** combining LLM-based accuracy, embedding-based semantic similarity, format adherence, and modality recognition, providing stable and informative feedback where traditional verifiable or MCQ-only rewards fall short.
+- We propose a **unified evaluation framework** for both text-only and image+text tasks using a Reference-based LLM-as-judge, capturing semantic correctness, reasoning, and contextual alignment.
+- Despite using only **~50K** instruction examples, MediX-R1 achieves state-of-the-art results across diverse medical LLM and VLM benchmarks, with particularly large gains on open-ended clinical tasks.
+
+---
+
+## Architecture
+
+<p align="center">
+  <img src="assets/medix-r1_arch.png" alt="MediX-R1 Architecture" width="100%">
+</p>
+
+---
+
+## Composite Reward Design
+
+MediX-R1 uses a multi-signal reward combining LLM-based accuracy, embedding-based semantic similarity, format adherence, and modality recognition. This stabilizes training and prevents reward hacking compared to single-signal approaches.
+
+<p align="center">
+  <img src="assets/reward_design_graph.png" alt="Reward Design" width="60%">
+</p>
+
+---
+
+## Qualitative Examples
+
+<p align="center">
+  <img src="assets/microscopy_qualitative.png" alt="Microscopy Example" width="85%">
+  <img src="assets/xray_qualitative.png" alt="X-ray Example" width="85%">
+</p>
+
+---
+
+## Training
+
+We provide training configs for all model sizes using GRPO and DAPO algorithms. The training pipeline uses a vLLM-based reward server for LLM-as-judge scoring during RL training.
+
+```bash
+cd training
+pip install -e .
+bash vllm_serve.sh       # Step 1: Start the reward server
+bash run_train.sh        # Step 2: Launch RL training
+bash merge_model.sh      # Step 3: Merge FSDP checkpoints
+```
+
+Training data: [MBZUAI/medix-rl-data](https://huggingface.co/datasets/MBZUAI/medix-rl-data) (~51K train, ~2.5K test samples)
+
+See [`training/README.md`](training/README.md) for detailed setup, configuration options, and per-model scripts.
+
+## Evaluation
+
+We propose a unified evaluation framework for both text-only (LLM) and image+text (VLM) tasks using a Reference-based LLM-as-judge across 17 medical benchmarks.
+
+```bash
+cd eval
+pip install uv && uv pip install -r requirements.txt
+bash eval.sh             # Run all phases: generate, evaluate, score
+```
+
+Supports self-hosted judge models via vLLM or [OpenRouter](https://openrouter.ai/) as a remote alternative. Results can be submitted to the [MediX Leaderboard](https://medix.cvmbzuai.com/leaderboard).
+
+See [`eval/README.md`](eval/README.md) for task selection, CLI reference, and MMMU-Medical evaluation.
+
+---
+
+## Model Zoo
+
+| Model | HuggingFace |
+|-------|-------------|
+| MediX-R1-2B | [MBZUAI/MediX-R1-2B](https://huggingface.co/MBZUAI/MediX-R1-2B) |
+| MediX-R1-8B | [MBZUAI/MediX-R1-8B](https://huggingface.co/MBZUAI/MediX-R1-8B) |
+| MediX-R1-30B | [MBZUAI/MediX-R1-30B](https://huggingface.co/MBZUAI/MediX-R1-30B) |
+
+---
+
+## Citation
+
+If you use MediX-R1 in your research, please cite our work as follows:
+
+```bibtex
+@misc{mullappilly2026medixr1openendedmedical,
+      title={MediX-R1: Open Ended Medical Reinforcement Learning}, 
+      author={Sahal Shaji Mullappilly and Mohammed Irfan Kurpath and Omair Mohamed and Mohamed Zidan and Fahad Khan and Salman Khan and Rao Anwer and Hisham Cholakkal},
+      year={2026},
+      eprint={2602.23363},
+      archivePrefix={arXiv},
+      primaryClass={cs.CV},
+      url={https://arxiv.org/abs/2602.23363}, 
+}
+```
+
+---
+
+## License
+
+This project is released for **research purposes only** under [*CC-BY-NC-SA 4.0*](https://creativecommons.org/licenses/by-nc-sa/4.0/legalcode.en) License. It is not intended for clinical or commercial use.
+
+Users are urged to employ MediX-R1 responsibly, especially when applying its outputs in real-world medical scenarios. It is imperative to verify the model's advice with qualified healthcare professionals and not rely on it for medical diagnoses or treatment decisions.
+
+---
+
+## Acknowledgements
+
+We are thankful to [EasyR1](https://github.com/hiyouga/EasyR1) (a fork of [veRL](https://github.com/volcengine/verl)) for their open-source RL training framework.
+
+This work was partially supported with  *NVIDIA Academic Grant 2025* and *MBZUAI-IITD* Research Collaboration Seed Grant. 
+
+We are grateful to [MBZUAI](https://mbzuai.ac.ae/) for compute and support.

mmproj-MediX-R1-8b-F16.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8d2b9604c5c483e7785003f630f99059f15ecd3db2592656296ed6e6f0bd1146
+size 1159030080