# Model: LLaMA (IFD Top 30%)

## 🔍 Purpose

Fine-tune `meta-llama/Llama-3.2-1B` on instruction samples with the **highest Instruction Flow Density (IFD)**.  
This group includes samples where the instruction contributes **least** to the model’s output (i.e., high IFD).

## 📂 Dataset

- `alpaca2000.csv`
  - IFD score 상위 30% (2000개 중 600개)
  - 기준: `PPL(y | x) / PPL(y)` (x: instruction+input, y: output)

## ⚙️ Training Config

- Model: `meta-llama/Llama-3.2-1B`
- Precision: `bf16` or `float32`
- Epochs: 3
- Max length: 2048
- Output: `output/llama_ifd`

## 🧪 Goal

Establish baseline performance of high-IFD samples, before splitting by instruction entropy.