---
license: apache-2.0
datasets:
- ai4bharat/samanantar
language:
- en
- hi
base_model:
- Helsinki-NLP/opus-mt-en-hi
---
# betterthan/opus-mt-en-hi

Fine-tuned MarianMT model for English → Hindi translation. This model is trained on **AI4Bharat's Samanantar dataset**, which contains over 10 million high-quality parallel sentences.

## 🔍 Model Details
- Base model: `Helsinki-NLP/opus-mt-en-hi`
- Fine-tuned on: `ai4bharat/samanantar` English–Hindi subset
- Total params: ~77M (MarianMT)
- Framework: Hugging Face Transformers

## 📊 Performance (BLEU / chrF on 500 samples from Namratap/En-Hindi)

| Domain            | Base BLEU | Fine-tuned BLEU | Base chrF | Fine-tuned chrF |
|-------------------|-----------|------------------|-----------|------------------|
| Healthcare        | 15.54     | **27.95**         | 38.06     | **54.09**         |
| Gen News          | 14.11     | **26.31**         | 39.07     | **52.98**         |
| Culture/Tourism   | 12.76     | **18.49**         | 35.07     | **41.32**         |
| Education         | 20.28     | **28.82**         | 43.84     | **49.68**         |

✅ BLEU improvements of **+8 to +13 points** across domains  
✅ chrF boosts up to **+16 points**, reflecting better fluency and coverage

## 🧠 Use Cases
- Book and news translation (Hindi)
- Offline/secure translation pipelines
- Domain-adapted fine-tuning

## 📁 Files Included
- `pytorch_model.bin` — fine-tuned model weights
- `config.json` — model architecture
- `tokenizer_config.json`, `vocab.json`, `source.spm`, `target.spm` — tokenizer
- `generation_config.json` — default decoding setup

## ⚖️ License
Apache 2.0 (Same as original model and Samanantar dataset)