---
license: mit
tags:
- game-ai
- supervised-fine-tuning
- game-of-21
- strategy-game
- distilgpt2
- peft
- lora
base_model: distilgpt2
---

# Game of 21 AI - SFT Model

## Model Description

This is the Supervised Fine-Tuned (SFT) model for the Game of 21 AI project. It was trained using LoRA (Low-Rank Adaptation) on optimal Game of 21 gameplay examples with distilgpt2 as the base model.

## Model Type: SFT (Supervised Fine-Tuning)

This model was trained to play the strategic Game of 21, where:
- Two players start at 0
- On each turn, a player adds 1, 2, 3, or 4 to the running total
- The player who reaches 21 or more loses
- The optimal strategy is to land on multiples of 5 (5, 10, 15, 20)

## Training Process

### SFT Training
This model was trained using supervised fine-tuning:
- **Training**: Supervised Fine-Tuning with LoRA
- **Dataset**: Training examples from optimal game play (debug mode uses subset)
- **Strategy**: Learned basic game strategy from examples
- **Performance**: ~40-60% win rate against optimal opponent
- **Training Time**: ~2-3 hours on Kaggle T4 GPU
- **Checkpoints**: Saved every 50 steps with auto-resume capability

## Performance Metrics

- **Win Rate**: ~40-60% against optimal opponent
- **Average Game Length**: ~4.2 turns
- **Invalid Moves**: Reduced invalid moves compared to baseline
- **Strategy**: Learned basic game strategy from training examples
- **Training**: Supervised Fine-Tuning with LoRA on game examples

## Usage

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

# Load base model
base_model = AutoModelForCausalLM.from_pretrained("distilgpt2")
tokenizer = AutoTokenizer.from_pretrained("distilgpt2")

# Load SFT adapter
model = PeftModel.from_pretrained(base_model, "araviiiman/game21-sft")
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

# Game prompt
prompt = """
Game Rules: Two players start at 0. On each turn, add 1, 2, 3, or 4 to the running total. 
The player who reaches 21 or more loses.

Current total: 5
What number (1, 2, 3, or 4) do you add?
I will add: """

# Generate response (using direct tokenization for distilgpt2 compatibility)
input_text = f"System: You are a strategic game player.\nUser: {prompt}"
input_ids = tokenizer(input_text, return_tensors="pt")
outputs = model.generate(input_ids.input_ids, max_new_tokens=5, temperature=0.7, do_sample=False)
response = tokenizer.decode(outputs[0][input_ids.input_ids.shape[1]:], skip_special_tokens=True)
print(response)
```

## Model Architecture

- **Base Model**: distilgpt2
- **Fine-tuning Method**: LoRA (Low-Rank Adaptation)
- **LoRA Parameters**: r=8, alpha=16, dropout=0.05
- **Target Modules**: c_attn, c_proj (distilgpt2 specific)
- **Parameters**: ~82M base model + adapter weights
- **Quantization**: 4-bit quantization for efficient inference

## Training Configuration

- **Learning Rate**: 5e-5
- **Batch Size**: 1 (effective: 8 with gradient accumulation)
- **Epochs**: 5
- **Sequence Length**: 64 (optimized for game prompts)
- **Optimizer**: adamw_torch
- **Mixed Precision**: FP16
- **Checkpoints**: Every 50 steps (max 3 saved)

## Training Data

The model was trained on examples of optimal Game of 21 play, including:
- Complete game transcripts with optimal moves
- Win/loss scenarios with proper strategy
- Strategic decision points at each game state
- Chat-formatted conversations for instruction following

## Limitations

- Model performance depends on the quality of training data
- May not perform optimally in all game scenarios
- Requires proper prompt formatting for best results
- Single epoch training may need more iterations for optimal performance

## Citation

```bibtex
@misc{game21-ai-sft,
  title={Game of 21 AI - SFT Model},
  author={Your Name},
  year={2025},
  url={https://huggingface.co/araviiiman/game21-sft}
}
```

## License

This model is licensed under the MIT License.