> ## Documentation Index
> Fetch the complete documentation index at: https://docs.monostate.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# What is AI Training

> Learn what this platform does and how it helps you

# Understanding AI Training

[![Ask DeepWiki](https://deepwiki.com/badge.svg)](https://deepwiki.com/monostate/aitraining)

AI Training lets you teach AI models to do specific tasks for your needs. Think of it like training a very smart assistant who learns from examples.

## Watch: Why Train Your Own Models?

<iframe width="100%" height="400" src="https://www.youtube.com/embed/JKfFNwQrLo4" title="Why Small Models Can Win" frameBorder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture" allowFullScreen />

## The Basic Idea

Imagine you want an AI that understands your company's products, or recognizes specific types of images, or writes in your brand's voice. Instead of using generic AI like ChatGPT, you can train your own specialized version.

## Training Methods Explained

<Table>
  <thead>
    <tr>
      <th>Method</th>
      <th>Plain English</th>
      <th>When to Use</th>
      <th>Technical Name</th>
    </tr>
  </thead>

  <tbody>
    <tr>
      <td>**Reward Training**</td>
      <td>Give points for good responses, remove points for bad ones</td>
      <td>Teaching AI to be helpful and safe</td>
      <td>PPO, RLHF</td>
    </tr>

    <tr>
      <td>**Preference Learning**</td>
      <td>Show AI examples of "good" vs "bad" outputs</td>
      <td>Making AI match human preferences</td>
      <td>DPO, ORPO</td>
    </tr>

    <tr>
      <td>**Knowledge Transfer**</td>
      <td>Copy abilities from expensive AI to cheaper ones</td>
      <td>Reducing costs while keeping quality</td>
      <td>Distillation</td>
    </tr>

    <tr>
      <td>**Example Learning**</td>
      <td>Show AI lots of examples of what you want</td>
      <td>General purpose training</td>
      <td>Fine-tuning, SFT</td>
    </tr>
  </tbody>
</Table>

## Making Training Affordable

Regular AI training needs expensive computers. We use tricks to make it work on normal hardware:

* **LoRA** - Only train small parts of the AI instead of everything
* **Quantization** - Use less precise numbers (like rounding 3.14159 to 3.14)
* **Flash Attention** - Skip unnecessary calculations
* **Gradient Checkpointing** - Trade speed for memory usage

Think of it like video compression - same quality, smaller file.

## What You Can Build

### For Business

* Customer service bots that know your products
* Document classifiers for your specific forms
* Sales email generators in your company voice

### For Research

* Species identifiers from photos
* Medical report analyzers
* Scientific paper summarizers

### For Fun

* Chatbots that talk like your favorite character
* Image classifiers for your hobbies
* Personal assistant that knows your preferences

## How Training Works

1. **Gather Examples** - Collect data showing what you want the AI to learn
2. **Pick a Starting Point** - Choose a pre-trained model (like Llama or GPT)
3. **Set Parameters** - Decide training speed, duration (or use our defaults)
4. **Train** - Let the computer learn from your examples
5. **Test** - Check if it works well
6. **Use** - Deploy your custom AI

## Supported Formats

### Input Data

* **Text Files** - .txt, .json, .jsonl for language models
* **Images** - .jpg, .png for vision models
* **CSV Files** - For tabular data (needs proper column structure)
* **Hugging Face Datasets** - Direct loading from the hub

### Output Models

* Standard PyTorch/Hugging Face formats
* Automatic LoRA adapter merging
* Push directly to Hugging Face Hub

## Key Features

What makes AITraining different:

* **Automatic Dataset Conversion** - Feed Alpaca, ShareGPT, Q\&A, or any format. We detect and convert automatically.
* **32 Chat Templates** - Llama, Gemma, Mistral, Qwen, and more. No manual formatting.
* **Hyperparameter Sweeps** - Optuna integration finds optimal settings automatically.
* **Built-in Evaluation** - BLEU, ROUGE, BERTScore, perplexity metrics during training.
* **Chat Interface** - Test your models immediately after training.

## Roadmap

Features we're working on:

* Model marketplace
* Cloud training with one click
* Advanced performance monitoring

Core training remains free and open source.

## Next Steps

<CardGroup cols={2}>
  <Card title="Installation" icon="download" href="/foundations/installation">
    Get AI Training running
  </Card>

  <Card title="Quick Start" icon="play" href="/foundations/quickstart">
    Train your first model
  </Card>
</CardGroup>

***

<Info>
  AI Training is a fork of Hugging Face's AutoTrain with additional training methods and optimizations. Free and open source.
</Info>