# GG Team Instruction-Tuned Adapters (LLaMA 3.2-3B)

This repository provides a collection of PEFT adapters (LoRA) trained on various instruction-tuning datasets using the base model **LLaMA 3.2-3B**. These adapters are developed by **GG Team - CSE476 @ Arizona State University**.

## Adapter Variants

| Folder | Dataset(s) Used | Description |
|--------|------------------|-------------|
| `llama-3.2-3B-sft` | Alpaca | Fine-tuned only on the original Alpaca dataset |
| `llama-3.2-3B-sft-dolly` | Alpaca + Dolly | Fine-tuned on Databricks' Dolly dataset |
| `llama-3.2-3B-sft-FLAN` | Alpaca + Dolly + FLAN | Fine-tuned on FLAN and Alpaca mixed |
| `sft_a_d` | Alpaca + Dolly | Combined dataset fine-tuning (Alpaca + Dolly) |
| `sft_a_d1` | Alpaca(cleaned) + Dolly | Combined dataset fine-tuning (Alpaca + Dolly) |

---

## 🛠️ Usage (with `peft`)

Here's an example of loading one of the adapters using 🤗 Transformers and PEFT:

```python
from peft import PeftModel
from transformers import AutoTokenizer, AutoModelForCausalLM

# Load base model
base_model = AutoModelForCausalLM.from_pretrained("meta-llama/Meta-Llama-3-3B")
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3-3B")

# Load adapter (choose one)
model = PeftModel.from_pretrained(base_model, "gg-cse476/gg/sft_a_d")

# Inference
prompt = "Explain how a rocket works in simple terms."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))