---
language:
  - en
  - sw
license: llama3.2
base_model: meta-llama/Llama-3.2-3B-Instruct
tags:
  - blood-donation
  - tanzania
  - swahili
  - bilingual
  - medical
  - unsloth
  - lora
---

# Blood Donation Advisory — Llama 3.2 3B (LoRA Adapters)

Bilingual **(English + Kiswahili)** blood donation advisor fine-tuned for Tanzania.

## Evaluation Results — Llama 3.2 3B

### NLP Quality (higher is better)

| Metric | Llama 3.2 1B | **Llama 3.2 3B** | Winner |
|---|:---:|:---:|:---:|
| ROUGE-1 | 0.4060 | **0.4410** | 3B ✅ |
| ROUGE-2 | 0.1845 | **0.2639** | 3B ✅ |
| ROUGE-L | 0.3039 | **0.3619** | 3B ✅ |
| BLEU | 0.2134 | **0.2461** | 3B ✅ |
| BERTScore Precision | 0.8741 | **0.8817** | 3B ✅ |
| BERTScore Recall | 0.8628 | **0.8676** | 3B ✅ |
| BERTScore F1 | 0.8683 | **0.8744** | 3B ✅ |
| Perplexity ↓ | 3.89 | **3.25** | 3B ✅ |

### Efficiency

| Metric | Llama 3.2 1B | Llama 3.2 3B |
|---|:---:|:---:|
| Training time (min) | **10.42** | 22.13 |
| Peak VRAM (GB) | **10.83** | 11.113 |
| LoRA VRAM (GB) | 9.738 | **8.361** |

## System Prompt
```
You are a helpful blood donation advisor serving Tanzania.
Always respond in the same language the user writes in.
```

## Quick Start (LoRA Adapters)
```python
from unsloth import FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="nsomazr/blood-donation-llama32-3b-lora",
    max_seq_length=8192, dtype=None, load_in_4bit=True,
)
```