| --- |
| language: |
| - he |
| - en |
| license: apache-2.0 |
| library_name: mamba |
| tags: |
| - mamba2 |
| - moe |
| - hebrew |
| - finance |
| - legal |
| - ssm |
| model_name: HEBATRON |
| base_model: nvidia/nemotron-3-nano-30b-base |
| pipeline_tag: text-generation |
| --- |
| |
| # 🛡️ HEBATRON: Hebrew-Specialized Mamba2-MoE |
|
|
| HEBATRON is a state-of-the-art, high-performance language model specialized for the Hebrew language[cite: 1]. Developed through a collaboration between **PwC Israel** and **MAFAT**, it introduces a unique hybrid architecture combining **Mamba2** and **Mixture-of-Experts (MoE)**[cite: 1]. |
|
|
| ## 🚀 Model Summary |
| HEBATRON is designed to handle the structural and morphological complexities of Hebrew while providing linear scaling for long-context tasks[cite: 1]. It is a localized and enhanced version of the **Nemotron-3-Nano-30B** framework, optimized for native-level reasoning in Hebrew and English[cite: 1]. |
|
|
| --- |
|
|
| ## 📂 Technical Specifications |
|
|
| | Feature | Specification | |
| | :--- | :--- | |
| | **Model Name** | HEBATRON[cite: 1] | |
| | **Architecture** | Hybrid **Mamba2** (SSM) + **Sparse MoE**[cite: 1] | |
| | **Total Parameters** | 30B[cite: 1] | |
| | **Active Parameters** | ~3B per token[cite: 1] | |
| | **Context Window** | 65,536 (64k) tokens[cite: 1] | |
| | **Hardware** | NVIDIA Blackwell (B300) & H200 GPUs[cite: 1] | |
| | **Precision** | FP8 Mixed-Precision[cite: 1] | |
|
|
| --- |
|
|
| ## 🧬 Training Curriculum |
| The model was trained using a three-phase **Curriculum Learning** strategy[cite: 1]: |
|
|
| 1. **Phase 1: Formal Foundation (75.5B tokens)** |
| Focused on high-quality, structured Hebrew (legal, academic, and literary texts) to establish core grammatical rules[cite: 1]. |
| 2. **Phase 2: Colloquial Expansion (3.36B tokens)** |
| Integration of social media, forums, and informal web data to handle slang and modern registers[cite: 1]. |
| 3. **Phase 3: Long-Context Extension (20.4B tokens)** |
| Fine-tuning on dense, long-form documents to stabilize the 64k context window[cite: 1]. |
|
|
| > **Alignment:** Supervised Fine-Tuning (SFT) was performed on **2 million samples**, including localized knowledge distillation and the **"Hebrew IFEval"** dataset[cite: 1]. |
|
|
| --- |
|
|
| ## 📊 Performance Evaluation |
|
|
| ### Hebrew Reasoning Benchmarks |
| * **SNLI (Semantic Reasoning):** 91.2% accuracy[cite: 1] |
| * **Israeli Trivia:** 72.1% (+14pt vs base)[cite: 1] |
| * **Hebrew Average Reasoning:** 73.8% (Surpassing DictaLM-3.0-Thinking)[cite: 1] |
| * **GSM8K (Math):** 83.3% accuracy in native Hebrew[cite: 1] |
|
|
| ### English Reasoning Benchmarks |
| * **Psychometric Psi (EN):** 91.6%[cite: 1] |
| * **English Reasoning Average:** 86.0%[cite: 1] |
|
|
| --- |
|
|
| ## 🎯 Intended Use & Limitations |
| * **Intended Use:** Advanced Hebrew document analysis, long-context summarization (legal/technical), and complex bilingual reasoning[cite: 1]. |
| * **Limitations:** Users should verify outputs for factual accuracy as with any Large Language Model[cite: 1]. |
|
|
| --- |
|
|
| ## 🤝 Credits |
| * **Developed by:** PwC Israel & MAFAT[cite: 1] |
| * **Technical Lead:** Sarel Weinberger (Co-founder, Binatna)[cite: 1] |
| * **Research Collaborators:** Shaltiel Shmidman (Dicta), Dan Revital (PwC Next)[cite: 1] |
