| --- |
| tags: |
| - sentinel-manifold |
| - quantization |
| - edge-deployment |
| - int4 |
| - pytorch |
| license: mit |
| --- |
| |
| # 🦴 Sentinel Tiny Text — Sentinel-INT4 |
|
|
| **Production-ready quantized model for edge deployment.** |
|
|
| This is the Sentinel-INT4 quantized variant of [Sentinel Tiny Text](https://huggingface.co/5dimension/sentinel-tiny-text). |
|
|
| --- |
|
|
| ## 📊 Quantization Spec (SDM-v1 §4) |
|
|
| | Property | Value | |
| |----------|-------| |
| | **Format** | Sentinel-INT4 | |
| | **Zero-Point (Z)** | C₁ = -0.007994021806 (attracting fixed point) | |
| | **Scale (S)** | max\|w\| · (1/e) | |
| | **Compression** | 8× | |
| | **Size** | ~6.4 MB | |
| | **Parent model** | [5dimension/sentinel-tiny-text](https://huggingface.co/5dimension/sentinel-tiny-text) | |
|
|
| --- |
|
|
| ## 🧬 Mathematical Foundation |
|
|
| The Sentinel quantization uses the attracting fixed point C₁ as the zero-point, minimizing rounding noise by clustering weights around the dynamical attractor. |
|
|
| ``` |
| Quantization: q = round((w - C₁) / S) |
| Dequantization: w = q · S + C₁ |
| Where S = max|w| · (1/e) |
| ``` |
|
|
| --- |
|
|
| ## 🚀 Deployment Targets |
|
|
| | Platform | Compatible | |
| |----------|-----------| |
| | Mobile (iOS/Android) | ✅ | |
| | IoT devices | ✅ | |
| | Microcontrollers (ARM Cortex-M) | ✅ | |
| | Browser (ONNX/WebGL) | ✅ | |
| | Edge TPU | ⚠️ (needs conversion) | |
|
|
| --- |
|
|
| ## 📦 Download |
|
|
| ```python |
| from huggingface_hub import hf_hub_download |
| |
| # Download quantized weights |
| weights = hf_hub_download(repo_id="5dimension/sentinel-tiny-text-int4", filename="model_int4.pt") |
| ``` |
|
|
| --- |
|
|
| ## 🔗 Links |
|
|
| - **Parent model**: [5dimension/sentinel-tiny-text](https://huggingface.co/5dimension/sentinel-tiny-text) |
| - **Main repo**: [sentinel-manifold-discoveries](https://huggingface.co/5dimension/sentinel-manifold-discoveries) |
| - **Sentinel Hub**: [Interactive demo](https://huggingface.co/spaces/5dimension/sentinel-hub) |
|
|
| --- |
|
|
| **License:** MIT | **Edge-ready. Theorem-backed.** 🦴 |
|
|
