# Training Pipeline

Fine-tune Nemotron Nano 12B v2 VL on AWS service documentation so the model has built-in knowledge of new/recent AWS services and features.

## Overview

- **Base model**: `nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16`
- **Method**: QLoRA (4-bit quantization + LoRA adapters)
- **Data**: AWS documentation converted to instruction-response pairs
- **Target**: Push fine-tuned model to HF Hub for deployment

## Directory Structure

```
training/
├── prepare_data.py      # Convert raw docs → training JSONL
├── train.py             # QLoRA fine-tuning script
├── push_to_hub.py       # Merge adapter + push to HF Hub
├── requirements.txt     # Training dependencies
└── data/
    ├── raw/             # Put AWS documentation markdown files here
    └── train.jsonl      # Generated training data (output of prepare_data.py)
```

## Steps

### 1. Gather AWS documentation

Add markdown files to `training/data/raw/`. One file per service or feature:

```
training/data/raw/
├── amazon-q.md
├── s3-express-one-zone.md
├── aurora-serverless-v2.md
├── bedrock.md
└── ...
```

### 2. Generate training data

```bash
python training/prepare_data.py
```

This reads the raw docs and creates `training/data/train.jsonl` with instruction-response pairs.

### 3. Train on AWS (EC2 with GPU)

```bash
pip install -r training/requirements.txt
python training/train.py
```

Recommended: `g5.2xlarge` (A10G, 24GB VRAM) or `p3.2xlarge` (V100, 16GB).

### 4. Push to Hugging Face Hub

```bash
python training/push_to_hub.py
```

### 5. Update app.py

Change `MODEL_ID` in `app.py` to point at your fine-tuned model.