|
|
--- |
|
|
base_model: unsloth/gemma-2-9b-bnb-4bit |
|
|
library_name: peft |
|
|
license: apache-2.0 |
|
|
language: |
|
|
- en |
|
|
- mr |
|
|
--- |
|
|
|
|
|
# Model Card for Gemma2 7B - English to Marathi Translation |
|
|
|
|
|
## Model Details |
|
|
|
|
|
### Model Description |
|
|
|
|
|
This model is a fine-tuned variant of **Unsloth's Gemma2 7B**, trained for high-quality English-to-Marathi translations. Built on a robust transformer architecture, the model handles complex translations, idiomatic expressions, and long-context paragraphs effectively. It is optimized for efficient inference using 4-bit quantization. |
|
|
|
|
|
- **Developed by:** Devavrat Samak |
|
|
- **Model type:** Causal Language Model, fine-tuned for translation tasks. |
|
|
- **Language(s) (NLP):** English (en), Marathi (mr) |
|
|
- **License:** Apache-2.0 |
|
|
- **Finetuned from model:** unsloth/gemma-2-9b-bnb-4bit |
|
|
|
|
|
### Model Sources |
|
|
|
|
|
- **Repository:** [https://github.com/Devsam2898/Gemma2-Marathi] |
|
|
|
|
|
## Uses |
|
|
|
|
|
### Direct Use |
|
|
|
|
|
The model can be directly used for English-to-Marathi translations, including handling long-context paragraphs, noisy inputs, and code-mixed sentences. |
|
|
|
|
|
### Downstream Use |
|
|
|
|
|
The model can be integrated into applications for: |
|
|
- Chatbots with multilingual support. |
|
|
- Translating historical texts for research. |
|
|
- Localization of content for Marathi-speaking audiences. |
|
|
|
|
|
### Out-of-Scope Use |
|
|
|
|
|
- The model is not designed for real-time, high-speed translation in latency-critical systems. |
|
|
- It may not generalize well for highly domain-specific jargon without additional fine-tuning. |
|
|
|
|
|
## Bias, Risks, and Limitations |
|
|
|
|
|
- The model's translations might occasionally lose nuance or context in culturally significant expressions. |
|
|
- Performance may degrade for noisy data or highly informal text. |
|
|
|
|
|
### Recommendations |
|
|
|
|
|
- Users should validate translations in sensitive domains to ensure accuracy. |
|
|
- Consider additional fine-tuning for domain-specific tasks. |
|
|
|
|
|
## How to Get Started with the Model |
|
|
|
|
|
```python |
|
|
from transformers import AutoTokenizer |
|
|
from unsloth import Gemma2 |
|
|
|
|
|
# Load model and tokenizer |
|
|
model = Gemma2.from_pretrained("unsloth/gemma-2-9b-bnb-4bit") |
|
|
tokenizer = AutoTokenizer.from_pretrained("unsloth/gemma-2-9b-bnb-4bit") |
|
|
|
|
|
# Input and inference |
|
|
input_text = "The golden age of the Peshwas brought cultural and political prosperity to Maharashtra." |
|
|
inputs = tokenizer(input_text, return_tensors="pt") |
|
|
outputs = model.generate(inputs["input_ids"], max_length=128) |
|
|
translated_text = tokenizer.decode(outputs[0], skip_special_tokens=True) |
|
|
print(translated_text) |
|
|
|
