| language: | |
| - bn | |
| - en | |
| library_name: transformers | |
| tags: | |
| - merge | |
| - mergekit | |
| - low-resource | |
| - mathematics | |
| base_model: | |
| - Qwen/Qwen2.5-1.5B | |
| - Qwen/Qwen2.5-1.5B-Instruct | |
| license: apache-2.0 | |
| # 🇧🇩 Polymath-1.5B-Bengali-Math | |
| This model is a merge of `Qwen/Qwen2.5-1.5B` and `Qwen/Qwen2.5-1.5B-Instruct` created using the **SLERP** method. | |
| ## 🧪 Research Goal | |
| To investigate the "Capacity Gap" in low-resource languages (like Bengali) when transferring mathematical reasoning capabilities without fine-tuning. | |
| ## 📊 Performance | |
| - **English Math Logic:** 60% Accuracy (Retained Logic) ✅ | |
| - **Bengali Math Logic:** 0% Accuracy (Capacity Constraint Revealed) ⚠️ | |
| ## 🛠️ Method | |
| Merged using [MergeKit](https://github.com/arcee-ai/mergekit) with the following config: | |
| - **Method:** SLERP | |
| - **Precision:** float16 | |
| - **Parameters:** 1.5 Billion | |
| *Created for research and educational purposes.* | |