Present the laboratory and its projects, Maverick as one of them

README.md

@@ -8,32 +8,33 @@ pinned: false
 ---
 
 The Extended Reality Laboratory (XRLab) at Manisa Celal Bayar University works on virtual, augmented and mixed
-reality, 3D virtual environments and educational technology, and on the AI models behind them. This page lists the
-models and tools we release.
+reality, 3D virtual environments and educational technology, and on the AI models that run inside them. This page
+collects what we release in the open.
 
-## Maverick-4B-Unity-XR-Agent
+[Website](https://xrlab.cbu.edu.tr) · [GitHub](https://github.com/mcbu-xrlab)
 
-<a href="https://huggingface.co/ErenAta00/Maverick-4B-Unity-XR-Agent"><img src="https://huggingface.co/spaces/ExtendedRealityLabMCBU/README/resolve/main/maverick.jpg" alt="Maverick-4B-Unity-XR-Agent" width="100%"></a>
+## Projects
+
+### Maverick: language control for Unity XR scenes
+
+<a href="https://huggingface.co/ErenAta00/Maverick-4B-Unity-XR-Agent-1.1"><img src="https://huggingface.co/ErenAta00/Maverick-4B-Unity-XR-Agent-1.1/resolve/main/assets/banner.jpg" alt="Maverick-4B-Unity-XR-Agent 1.1" width="55%"></a>
 
 A 4B language model that turns spoken or typed English into actions in Unity XR scenes. It picks up, places and moves
 objects, switches devices on and off, moves the user between rooms, and asks which object you mean when a command is
-unclear. It runs offline through llama.cpp and needs about 3 GB of GPU memory.
-
-On 499 instructions written by people for the ALFRED benchmark, it chose the right action on the right object 83.8% of
-the time, ahead of models with up to 120B parameters. The [article](https://huggingface.co/blog/ErenAta00/maverick-4b-unity-xr-agent) explains how it was trained and tested.
+unclear. It runs offline through llama.cpp on a 4 GB laptop GPU. Version 1.1 adds undo and corrections, shared
+attention for several people, actions across rooms and short plans.
 
-Version 1.1 adds undo and corrections, shared attention for several people, actions across rooms and short plans: [Maverick-4B-Unity-XR-Agent 1.1](https://huggingface.co/ErenAta00/Maverick-4B-Unity-XR-Agent-1.1).
+On 499 instructions written by people for the ALFRED benchmark, 1.1 chose the right action on the right object 83.2% of
+the time (1.0: 83.8%), ahead of the other models we tested, which scored up to 67.1%.
 
-## Releases
+- [Maverick-4B-Unity-XR-Agent 1.1](https://huggingface.co/ErenAta00/Maverick-4B-Unity-XR-Agent-1.1), the current model
+- [Maverick-4B-Unity-XR-Agent 1.0](https://huggingface.co/ErenAta00/Maverick-4B-Unity-XR-Agent), the first release
+- [Maverick XR Agent for Unity](https://github.com/mcbu-xrlab/Maverick-Unity), the Unity package that connects the model to your scene ([Hugging Face mirror](https://huggingface.co/ErenAta00/Maverick-Unity))
+- [Offline Voice Control for Unity XR Scenes](https://huggingface.co/blog/ErenAta00/maverick-4b-unity-xr-agent), the article on how 1.0 was trained and tested
 
-| Release | Type | Description |
-|---|---|---|
-| [Maverick-4B-Unity-XR-Agent 1.1](https://huggingface.co/ErenAta00/Maverick-4B-Unity-XR-Agent-1.1) | Model | Adds undo and corrections, shared attention, actions across rooms and short plans |
-| [Maverick-4B-Unity-XR-Agent](https://huggingface.co/ErenAta00/Maverick-4B-Unity-XR-Agent) (1.0) | Model | Tool calling for Unity XR scenes, GGUF |
-| [Maverick XR Agent for Unity](https://github.com/mcbu-xrlab/Maverick-Unity) | Unity package | Connects the model to your scene; installs from the Package Manager ([Hugging Face mirror](https://huggingface.co/ErenAta00/Maverick-Unity)) |
-| [Offline Voice Control for Unity XR Scenes](https://huggingface.co/blog/ErenAta00/maverick-4b-unity-xr-agent) | Article | How Maverick was trained and tested |
-| [XRLab-SDK](https://github.com/mcbu-xrlab/XRLab-SDK) | Python SDK | Shows which vision-language models fit on an XR headset and what they cost to run |
+### XRLab-SDK: vision-language models on XR headsets
 
-## Links
+A Python SDK that reports what a headset is, which vision-language model configurations fit on it, and what running
+them costs. It is built from a study of about 10,000 inferences on a Meta Quest 3.
 
-[Website](https://xrlab.mcbu.edu.tr) · [GitHub](https://github.com/mcbu-xrlab)
+- [XRLab-SDK](https://github.com/mcbu-xrlab/XRLab-SDK) on GitHub