| --- |
| license: apache-2.0 |
| base_model: VTXAI/vtx-jev-mod-multilingual |
| pipeline_tag: feature-extraction |
| language: |
| - multilingual |
| - en |
| - hi |
| - bn |
| - ta |
| - te |
| - mr |
| - ur |
| - ar |
| - es |
| - ru |
| - zh |
| tags: |
| - sentence-transformers |
| - static-embeddings |
| - model2vec |
| - moderation |
| - safety |
| - abuse-detection |
| - lf2 |
| - 2bit-quantization |
| - cpu-optimized |
| --- |
| |
| # VTX-MOD-1 |
|
|
| **VTX-MOD-1** is an ultra-fast, lightweight **Multilingual Content Moderation and Safety Gate** distilled from `VTXAI/vtx-jev-mod-multilingual` (`mmBERT-base` foundation) into a **256-dimensional static embedding architecture** with native **LF2 2-bit quantization**. |
|
|
| It is designed for real-time abuse detection, vulgarity filtering, and toxicity screening across **100+ languages**, including Indian scripts (Hindi, Bengali, Tamil, Telugu, Marathi, Urdu) and international languages (Spanish, Russian, Arabic, French, German, Chinese, English). |
|
|
| --- |
|
|
| ### Highlights |
|
|
| - ⚡ **Ultra-Low Latency**: Single text classification in **0.06 ms** (LF2 2-bit) to **0.28 ms** (FP32) on standard CPU. |
| - 🚀 **High Throughput**: Processes over **41,000 queries per second** on CPU. |
| - 🗜️ **2-Bit LF2 Quantization**: |
| - Compresses from **124.9 MB down to 19.5 MB** in RAM and on disk (over **6.4x reduction**). |
| - Uses native blockwise integer quantization (4 weights per `uint8` byte). |
| - 🌐 **100+ Languages Coverage**: 256,000 multilingual vocabulary preserves full native words across all Indian and international scripts without fragmentation. |
| - 📦 **Dual Framework Support**: Directly loadable via `SentenceTransformer`, `model2vec`, or the bundled `inference.py`. |
|
|
| --- |
|
|
| ### Quick Start: Content Moderation & Toxicity Scoring |
|
|
| Using native anchor vector matching: |
|
|
| ```python |
| import numpy as np |
| from model2vec import StaticModel |
| |
| # Load model (requires only ~19.5 MB RAM in LF2 or ~125 MB in FP32) |
| model = StaticModel.from_pretrained("VTXAI/VTX-MOD-1") |
| |
| def normalize(v): |
| return v / (np.linalg.norm(v, axis=-1, keepdims=True) + 1e-9) |
| |
| # Pre-compute reference safe and toxic concept vectors |
| safe_anchor = "polite safe respectful helpful friendly conversation धन्यवाद शुक्रिया thank you" |
| toxic_anchor = "offensive abuse toxic hate vulgar insult गाली बकवास হারামজাদা idiot stupid fool" |
| anchors = normalize(model.encode([safe_anchor, toxic_anchor])) |
| safe_vec, toxic_vec = anchors[0], anchors[1] |
| |
| def check_moderation(text: str) -> dict: |
| v = normalize(model.encode([text])[0]) |
| sim_toxic = float(np.dot(v, toxic_vec)) |
| sim_safe = float(np.dot(v, safe_vec)) |
| prob_toxic = 1.0 / (1.0 + np.exp(-(sim_toxic - sim_safe) * 8.0)) |
| return { |
| "text": text, |
| "is_harmful": prob_toxic >= 0.5, |
| "toxic_probability": round(prob_toxic, 4), |
| } |
| |
| # Test cases |
| print(check_moderation("तू बिल्कुल बेवकूफ और बकवास इंसान है।")) |
| # {'is_harmful': True, 'toxic_probability': 0.844} |
| |
| print(check_moderation("नमस्ते, क्या आप ट्रांसफॉर्मर मॉडल कैसे काम करता है समझा सकते हैं?")) |
| # {'is_harmful': False, 'toxic_probability': 0.254} |
| |
| print(check_moderation("You are an absolute idiot and completely worthless.")) |
| # {'is_harmful': True, 'toxic_probability': 0.859} |
| ``` |
|
|
| --- |
|
|
| ### Usage with Sentence Transformers |
|
|
| ```python |
| from sentence_transformers import SentenceTransformer |
| |
| model = SentenceTransformer("VTXAI/VTX-MOD-1") |
| embeddings = model.encode([ |
| "This is a respectful message.", |
| "chup kar saale bhadwe pagal", |
| "வணக்கம் நண்பா" |
| ]) |
| print(embeddings.shape) # (3, 256) |
| ``` |
|
|
| --- |
|
|
| ### Performance & Latency Benchmarks (CPU) |
|
|
| | Benchmark | FP32 Static (`model.safetensors`) | LF2 2-Bit (`model_lf2.safetensors`) | |
| | :--- | :--- | :--- | |
| | **Model Size (RAM / Disk)** | 124.9 MB | **19.5 MB (6.4x compression)** | |
| | **Single Query Latency** | 0.28 ms | **0.06 ms** (60 microseconds) | |
| | **Batch Throughput (Batch=20)** | 8,637 queries/sec | **41,720 queries/sec** | |
| | **Multilingual Native Accuracy** | **88.9% – 100%** | **88.9% – 100%** | |
|
|
