docs: update README to enhance model overview, configuration, and deployment details

README.md

@@ -23,57 +23,82 @@ base_model:
 </p>
 
 ## 📖 Overview
-**TinyMyo** is a lightweight, Transformer-based foundation model designed specifically for surface electromyography (sEMG) signal processing. Unlike large-scale models, the TinyMyo family (including the 3.6M parameter base model and the ultra-compact 1.9M parameter **TinyissimoMyo**) is purpose-built for **ultra-low-power edge deployment**. It enables real-time motor intent decoding, neuromuscular assessment, and human-machine interaction directly on microcontrollers like the GAP9.
-
-## 🚀 Key Highlights
-*   **Generalist Foundation:** Pre-trained on a massive, heterogeneous corpus of >480 GB of EMG data (NinaPro DB6/7, EMG2Pose) using self-supervised masked reconstruction.
-*   **Edge-Ready:** The first EMG foundation model demonstrated on an ultra-low-power MCU (GAP9), achieving sub-100ms inference for real-time applications.
-*   **Highly Efficient:** Just 3.6M parameters (1.9M for TinyissimoMyo), ensuring low latency and high energy efficiency (~45 mJ per inference).
-*   **Versatile:** Achieves state-of-the-art (SoA) performance across hand gesture classification, kinematic regression, and speech processing.
+**TinyMyo** is a lightweight, Transformer-based foundation model designed specifically for surface electromyography (sEMG) signal processing. Unlike large-scale models, the TinyMyo family (including the 3.6M parameter base model and the ultra-compact 1.9M parameter **TinyssimoMyo**) is purpose-built for **ultra-low-power edge deployment**. It enables real-time motor intent decoding, neuromuscular assessment, and human-machine interaction directly on microcontrollers like the GAP9.
+
+## ⚙️ Model Configuration
+*   **Base model:** 8-layer bidirectional Transformer encoder.
+*   **TinyssimoMyo:** 4-layer compact variant.
+*   **Embedding dimension:** 192.
+*   **Attention heads:** 3.
+*   **Temporal patch size:** 20 samples.
+*   **Default input:** 16 channels × 1000 samples.
+*   **Maximum channels:** 16.
+*   **Tokenization:** Channel-independent patching with 50 temporal patches per channel and 800 tokens for the default input.
+*   **Position encoding:** Rotary Position Embeddings (RoPE), with temporal positions reset for each channel.
+*   **Training objective:** Self-supervised masked patch reconstruction.
 
 ## 🧠 Model Architecture
-*   **Core:** 8-layer bidirectional Transformer encoder (4-layer for TinyissimoMyo).
-*   **Embeddings:** 192-dimensional latent space with 3 attention heads.
-*   **Tokenization:** Channel-independent patching (20 samples per patch) utilizing Rotary Position Embeddings (RoPE) to preserve temporal alignment across channels without spurious cross-channel ordering.
-*   **Deployment:** Optimized via offline liveness analysis, multi-level memory tiling, and INT8 fixed-point quantization for resource-constrained hardware execution.
-
-## 📊 Performance Benchmarks
-
-| Task | Dataset | Metric | TinyMyo Result |
-| :--- | :--- | :--- | :--- |
-| **Gesture Classification** | NinaPro DB5 | Accuracy | **87.98%** |
-| **Gesture Classification** | EPN-612 | Accuracy | **96.57%** |
-| **Gesture Classification** | UCI EMG | Accuracy | **97.10%** |
-| **Gesture Classification** | Generic Neuromotor Interface | CLER | **0.142** |
-| **Kinematic Regression** | NinaPro DB8 | MAE | **8.8°** |
-| **Speech Synthesis** | Gaddy | WER | **33.54%** |
-| **Speech Recognition** | Gaddy | WER | **33.95%** |
+TinyMyo uses channel-independent patch embeddings followed by a bidirectional
+Transformer encoder. Tokens are ordered channel-major, while RoPE positions reset
+for each channel so flattening does not introduce a false temporal distance between
+channels. The learned channel embedding identifies a channel slot; it does not
+encode physical electrode coordinates.
+
+For deployment, the family can be paired with offline liveness analysis,
+multi-level memory tiling, and INT8 fixed-point execution. See the paper and model
+card for deployment measurements.
 
 ## ⚡ Deployment (GAP9 MCU)
-TinyMyo bridges the gap between high-performance deep learning and stringent wearable constraints. We provide two variants to balance the accuracy-latency trade-off:
+TinyMyo is designed for resource-constrained deployment on platforms such as the
+GAP9 MCU. The repository does not treat deployment measurements as model
+configuration; see the [paper](https://arxiv.org/abs/2512.15729) and the
+[Hugging Face model card](https://huggingface.co/PulpBio/TinyMyo) for current results.
 
 ### TinyMyo (3.6M Parameters)
-*   **Inference Time (5s window):** 0.785 s
-*   **Energy Consumption:** 44.91 mJ
-*   **Power Envelope:** 57.18 mW
+*   8-layer encoder configuration.
 
-### TinyissimoMyo (1.9M Parameters)
-*   **Inference Time (5s window):** 0.496 s
-*   **Inference Time (1s window):** **0.089 s** *(Sub-100ms regime, ideal for real-time prosthetic control)*
+### TinyssimoMyo (1.9M Parameters)
+*   4-layer compact configuration.
 
 ## 🛠️ Getting Started
-TinyMyo is part of the[BioFoundation](https://github.com/pulp-bio/BioFoundation) ecosystem.
+TinyMyo is part of the [BioFoundation](https://github.com/pulp-bio/BioFoundation) ecosystem.
 
 ### Prerequisites
 Install the required dependencies from the [BioFoundation repository](https://github.com/pulp-bio/BioFoundation).
 
 ### Loading & Fine-tuning
-You can easily fine-tune the pre-trained weights for your specific task:
+Fine-tune a pretrained checkpoint using the BioFoundation training entry point:
 
 ```bash
 python run_train.py +experiment=TinyMyo_finetune pretrained_safetensors_path={*.safetensors}
 ```
 
+Available checkpoint groups include pretraining, DB5, EPN-612, UCI EMG, and DB8.
+Use the matching configuration and checkpoint for each task. The released
+classification labels are:
+
+| Dataset | Input channels | Classes | Label convention | Checkpoint |
+| :--- | ---: | ---: | :--- | :--- |
+| NinaPro DB5 | 16 | 53 | 52 gestures + resting | `DB5/DB5_finetune_5sec.safetensors` |
+| EPN-612 | 8 | 6 | 5 gestures + hand relaxed | `EPN612/EPN_finetune_5sec.safetensors` |
+| UCI EMG | 8 | 6 | Dataset-specific six-class gesture labels | `UCI_EMG/UCI_finetune_5sec.safetensors` |
+| NinaPro DB8 | 16 | 5 | Regression outputs | `DB8/DB8_finetune_500ms.safetensors` |
+
+The model configuration must match the checkpoint, especially `in_chans`,
+`num_classes`, and `task`. The built-in BioFoundation `TinyMyo_finetune`
+experiment currently points to UCI EMG data and can be launched from a
+BioFoundation checkout with:
+
+```bash
+python -u run_train.py +experiment=TinyMyo_finetune \
+  model.in_chans=8 \
+  pretrained_safetensors_path=/absolute/path/to/TinyMyo/UCI_EMG/UCI_finetune_5sec.safetensors
+```
+
+DB5, EPN-612, and DB8 require matching dataset-specific data-module settings
+in BioFoundation before training. Benchmark results
+and the latest experimental protocols are maintained in the paper and model card.
+
 ## 📜 License & Citation
 This model is licensed under **CC BY-ND 4.0**. If you find TinyMyo useful in your research, please cite our paper:
 
@@ -84,6 +109,7 @@ This model is licensed under **CC BY-ND 4.0**. If you find TinyMyo useful in you
       year={2026},
       eprint={2512.15729},
       archivePrefix={arXiv},
-      primaryClass={eess.SP}
+      primaryClass={eess.SP},
+      url={https://arxiv.org/abs/2512.15729},
 }
-```
+```