---
language: ta
license: llama3
tags:
- tamil
- government-schemes
- qlora
- unsloth
- llama3
- instruction-tuning
base_model:
- unsloth/Llama-3.2-1B-Instruct-bnb-4bit
datasets:
- MURUGAYESU/tamil-govt-scheme-qa
metrics:
- rouge
- bertscore
---

# Tamil Government Scheme QA — Llama-3.2-1B-Instruct-bnb-4bit

Fine-tuned on Tamil-language QA pairs covering 10 major Indian and Tamil Nadu
government welfare schemes (PM-KISAN, Amma Unavagam, MGNREGS, Ayushman Bharat, etc.)

## Training details
- **Base model**: meta-llama/Meta-Llama-3-8B
- **Method**: QLoRA — 4-bit NF4 + LoRA (r=16, alpha=32, all projections)
- **Framework**: Unsloth + TRL SFTTrainer
- **Dataset**: 270 train / 30 val (Tamil script, deduplicated)
- **ROUGE-L**: 0.0000 | **BERTScore F1**: 0.7482
- **Training time**: ~125 min on Colab T4

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained("MURUGAYESU/tamil-govt-scheme-llama3-8b", torch_dtype=torch.float16, device_map='auto')
tokenizer = AutoTokenizer.from_pretrained("MURUGAYESU/tamil-govt-scheme-llama3-8b")

prompt = '''### Instruction:
கீழே கொடுக்கப்பட்ட அரசு திட்டம் பற்றிய கேள்விக்கு தமிழில் விரிவான பதிலளிக்கவும்.

### Input:
PM-KISAN திட்டத்திற்கு யார் தகுதியானவர்கள்?

### Response:
'''
inputs = tokenizer(prompt, return_tensors='pt').to('cuda')
output = model.generate(**inputs, max_new_tokens=150, temperature=0.1, do_sample=True)
print(tokenizer.decode(output[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True))
```

## Covered schemes
PM-KISAN, Amma Unavagam, MGNREGS, Pradhan Mantri Awas Yojana, Ayushman Bharat,
Kisan Credit Card, Sukanya Samridhi, CM Solar Scheme TN,
Mukhyamantri Kalaignar Scheme, National Scholarship Portal

## Disclaimer
This model provides general educational information only. Always verify with official
government portals (india.gov.in, tn.gov.in) before making any decisions.