---
license: apache-2.0
language:
- en
- ta
---


# Tamil-Qwen3-4B-Inst

**Tamil-Qwen3-4B-Inst** is a lightweight Tamil-English bilingual instruction-tuned model designed for efficient deployment and strong performance on instruction-following tasks.

---

## Model Overview

We developed a specialized model by adapting state-of-the-art open-source base models through:

- **Continual Pretraining** on the **Tamil Wikipedia** dataset  
- **Instruction Finetuning** using high-quality, human-annotated Tamil instruction datasets from the **Aya Dataset**

---
## Model Summary

| Feature | Description |
|----------|-------------|
| **Base Model** |Qwen3-4B(trained in tamil wikipedia dataset) |
| **Parameters** | 4B |
| **Datasets** | https://huggingface.co/datasets/CohereLabs/aya_dataset,https://huggingface.co/datasets/wikimedia/wikipedia/viewer/20231101.ta |
| **Training Precision** | bfloat16 |
| **Epochs (Total)** | 6|
| **Languages** | Tamil, English |
| **Use Case** | Instruction following, conversational AI, and Tamil language tasks |

---
## Prompting Format

**Prompt Template Without Input**

```
{system_prompt}
### Instruction:
{instruction or query}
### Response:
{response}
```

**Prompt Template With Input**

```
{system_prompt}
### Instruction:
{instruction or query}
### Input:
{input}
### Response:
{response}
```

## Citation

If you use this model in your research, please cite:

```bibtex
@misc{tamilqwen3_4b_inst,
  title={Tamil-Qwen3-4B-Inst: Efficient Bilingual Instruction-Tuned Model},
  author={AITamilNadu},
  year={2025},
  url={https://huggingface.co/aitamilnadu/Tamil-Qwen3-4B-Inst}
}

```

## Evaluation and Benchmarks

| Benchmark | Score |
|------------|-------|
| **Average** | 52.08% |
| ARC Challenge | 45.48% |
| HellaSwag | 61.64% |
| MMLU | 56.05% |
| TruthfulQA | 39.58% |
| Winogrande | 59.43% |
| GSM8K | 40.64% |