# Sumobot AI Assistant Models

This repository contains three Jupyter notebooks for training and exporting AI models for Sumobot gameplay:

- `train_ml.ipynb`: A multi-class classification model (ML) using TensorFlow
- `train_slm.ipynb`: A sequence-level model (SLM) using a GPT-style transformer
- `train_llm.ipynb`: A large language model (LLM) using LoRA fine-tuning

---

## Get Started

### Clone including the datasets

Note: This will fetch all dataset/*.csv sumobot logs

1. Run `git lfs install`
2. Run `git clone https://huggingface.co/datasets/arbyazra123/sumobot_ml`

### Clone without datasets

Note: By running the .ipynb files, it will scan local `dataset` folder. If it's empty, will fetch from online

1. Run `GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/datasets/arbyazra123/sumobot_ml`

### Running Online Notebooks

1. Upload ipynb file that you want to run, for example `train_llm.ipynb`
2. Upload `dataset_helper.py` to the notebook, for example in google colab, drag and drop `dataset_helper.py` on colab's Files
3. `train_llm.ipynb` requires to install `trl`. Therefore, install the package by running `!pip install trl`
4. `train_ml.ipynb` requires to install `tensorflow`. Therefore, install the package by running `!pip install tensorflow`
5. If online notebook still raising `ModuleNotFound`, upload `requirements.txt` in online notebooks, then run `!pip install -r requirements.txt`
6. It's done, you dont have to follow next steps.

### Running the Offline Notebooks

#### Installing Requirements

Install all dependencies using `pip` with Command Prompt / Powershell (Windows) or Terminal (Macos/Linux):

```bash
pip install -r requirements.txt
```

> This installs all required packages (PyTorch, TensorFlow, Hugging Face, ONNX, etc.) and works on **Windows, macOS, and Linux**.

---

#### Prerequisites
- [Jupyter Notebook](https://jupyter.org/install) installed
- Python 3.10+
- Dataset will be fetched from HugginFace

---

#### 1. Run `train_ml.ipynb` (Multi-Label Classification)

1. Open Jupyter:
   ```bash
   jupyter notebook
   ```
2. Navigate to `train_ml.ipynb`
3. Run all cells

> This trains a Keras model and exports it as `ml.onnx`

Output: `ml.onnx` and `action_labels.json`

---

#### 2. Run `train_slm.ipynb` (Sequence Model)

1. Open Jupyter
2. Navigate to `train_slm.ipynb`
3. Run all cells

> This trains a GPT-style SLM and exports as `slm.onnx`

Output: `slm.onnx` and `slm_tokenizer.json`

> ⚠️ **Note**: This model works best with short, structured text inputs (e.g., bot positions, angles). You can adjust `max_dataset` to limit training data.

---

#### 3. Run `train_llm.ipynb` (LLM with LoRA)

1. Open Jupyter
2. Navigate to `train_llm.ipynb`
3. Run all cells

> This fine-tunes a Qwen-0.5B model using LoRA and exports the adapter

Output: `adapters/qwen2.5_0.5b_lora/` (contains model + tokenizer)

> Requires ~4-8GB VRAM for full training. For low-end systems, reduce `max_dataset` or `num_train_epoch`.

---

## Adjusting Configurations

### Modify key settings in each notebook:

#### In `train_ml.ipynb`:
- Change `features` to include/exclude bot or enemy stats
- Adjust `batch_size`, `epochs`, or `learning_rate` for performance
- Set `max_dataset` to limit training (e.g., `max_dataset=1000`)

#### In `train_slm.ipynb`:
- Adjust `block_size`, `batch_size`, `n_layers`, `lr`
- Set `max_dataset` to control training data size
- Change `prompt` in `generate()` to test different inputs

#### In `train_llm.ipynb`:
- Set `max_dataset` to limit training (e.g., `max_dataset=10000`)
- Adjust `batches_per_device`, `gradient_accumulation`, `learning_rate`, or `num_train_epoch`
- Modify `prompt` in inference section to test different behaviors

---

## Notes for All Platforms

| Platform | Notes |
|--------|-------|
| **Windows** | Use `pip install` in terminal; Jupyter works via `jupyter notebook` or `jupyter lab` |
| **macOS** | Use `pip` or `pip3`; Jupyter works via `jupyter notebook` or `jupyter lab` |
| **Linux (Ubuntu/Debian)** | Use `pip install` and `jupyter notebook` |

> To rerun: Restart Jupyter kernel and re-run from the top.

---

## Known Issues & Tips

- `train_llm.ipynb`: may fail on CPU due to lack of VRAM — use `device_map="auto"` or limit `max_dataset`
- `train_ml.ipynb`: Ensure `cleaned_log.csv` for  has columns: `Name`, `Duration`, and the listed features
- Test outputs by changing the prompt or input data

---

## Example Prompt Test (for all models)

For `train_slm.ipynb` and `train_llm.ipynb`, you can test with:

```text
BotPos=[2.23,2.25], BotRot=228, EnemyPos=[2.87,0.39], EnemyRot=87, AngleToEnemy=-29.68, AngleToEnemyScore=0.87, DistanceToEnemyScore=0.79, NearBorderArenaScore=0.42, FacingToArena=0.65. Suggested Action:
```

> The models will generate a response based on the provided game state.

---

## Summary

| Model | Use Case | Exported To |
|------|---------|-------------|
| ML | Action classification | `ml.onnx` |
| SLM | Sequence prediction | `slm.onnx` |
| LLM | Natural language reasoning | `adapters/qwen2.5_0.5b_lora/` |
```