Upload README.md with huggingface_hub

README.md

@@ -1,10 +1,80 @@
 ---
 title: README
-emoji: 🐠
+emoji: 📱
 colorFrom: yellow
-colorTo: red
+colorTo: blue
 sdk: static
 pinned: false
 ---
 
-Edit this `README.md` markdown file to author your organization card.
+![MobileTransformers](https://huggingface.co/spaces/mobiletransformers/README/resolve/main/mobiletransformers_banner.png)
+
+# MobileTransformers
+
+**Export a Hugging Face model, pull it onto a phone, then chat with it, retrieve over your own
+documents, classify text, fine-tune it, merge the adapter into the weights, and let it call tools —
+entirely on the device.** No server, no inference API, no data leaving the phone.
+
+Built on **ONNX Runtime**, with hardware-accelerated execution providers (XNNPACK, NNAPI, QNN) for
+both inference *and* training on Android.
+
+👉 **[github.com/martinkorelic/mobiletransformers](https://github.com/martinkorelic/mobiletransformers)**
+
+## The packages here are not plain models
+
+Every repo in this organization holds an **exported MobileTransformers package**: a manifest, ONNX
+stages per variant, and a weight-handoff map that lets a trained adapter be merged into the inference
+weights *on the phone*. `transformers`, `optimum` and plain `onnxruntime` cannot load them — the
+Android SDK is what reads them.
+
+```kotlin
+val model = MobileTransformers.fromPretrained(
+    context = context,
+    repoId  = "mobiletransformers/SmolLM2-135M-Instruct",
+)
+```
+
+Every package ships **both an inference and a training stage**, because on-device fine-tuning is the
+point of the project — a package you cannot train demonstrates half of it.
+
+## The shelf
+
+| model | task | download | fine-tuning | good for |
+| --- | --- | --- | --- | --- |
+| [SmolLM2-135M-Instruct](https://huggingface.co/mobiletransformers/SmolLM2-135M-Instruct) | text-generation | 663 MB | LoRA | **start here** — fastest to pull and to train |
+| [gemma-3-270m-it](https://huggingface.co/mobiletransformers/gemma-3-270m-it) | text-generation | 1814 MB | **MARS** | the shared-adapter method this project researches |
+| [functiongemma-270m-it](https://huggingface.co/mobiletransformers/functiongemma-270m-it) | text-generation | 3557 MB | LoRA | tool / function calling |
+| [Qwen2.5-0.5B-Instruct](https://huggingface.co/mobiletransformers/Qwen2.5-0.5B-Instruct) | text-generation | 2554 MB | LoRA | chat quality, and the memory ceiling |
+| [all-MiniLM-L6-v2](https://huggingface.co/mobiletransformers/all-MiniLM-L6-v2) | text-classification | 94 MB | LoRA | retrieval, and the smallest training run |
+| [distilbert-sst2-english](https://huggingface.co/mobiletransformers/distilbert-sst2-english) | text-classification | 270 MB | LoRA | classification with real trained labels |
+
+Sizes are the **inference group** — what a plain install downloads — measured from each package's
+manifest. Asking for the training or retrieval group adds to it.
+
+> Some repos are currently private while the project is being prepared for release. Set an `HF_TOKEN`
+> with access, or start with the public ones.
+
+## MARS
+
+**Multi-Adapter Rank Sharing** is this organization's own fine-tuning method. Instead of every layer
+carrying its own low-rank adapter, layers share a down-projection — so the trainable parameter count
+grows with *rank* rather than with *depth*. On
+[gemma-3-270m-it](https://huggingface.co/mobiletransformers/gemma-3-270m-it) that is **279,936
+trainable parameters against a 268M backbone**, small enough to train in a phone's memory budget.
+
+The research behind it, including the on-device training methodology and the experimental results, is
+in the thesis: [*Parameter-Efficient Tuning of Large Language Models on Mobile
+Devices*](https://repozitorij.uni-lj.si/IzpisGradiva.php?lang=eng&id=175561).
+
+## Citation
+
+If you are using this framework for your own work, please cite:
+
+```bibtex
+@misc{mobiletransformers2025,
+  author       = {Koreli\v{c}, Martin and Pejovi{\'c}, Veljko},
+  title        = {MobileTransformers: An On-Device LLM PEFT Framework for Fine-Tuning and Inference},
+  year         = {2025},
+  howpublished = {\url{https://gitlab.fri.uni-lj.si/lrk/mobiletransformers}}
+}
+```