🎮 HIDER Agent - Fine-tuned Gemma 3 270M

Fine-tune Gemma 3 270M to play Hide-and-Seek using Modal (serverless GPU platform with free credits)!

🚀 Quick Start

# 1. Install Modal
pip install modal

# 2. Setup Modal account (get free credits!)
modal setup

# 3. Train the model on cloud GPU
modal run train_modal.py

# 4. Test the model
modal run test_model.py

That's it! Training takes ~20 minutes on a free T4 GPU.

📖 What is this?

This project fine-tunes Google's Gemma 3 270M model to control a HIDER agent in a Hide-and-Seek game. The model learns to:

• Process 112-dimensional observation vectors
• Output optimal actions: [move_x, move_y, move_z, lock, grab]
• Make strategic decisions in 3D space

🎯 Features

• ✅ Modal Integration: Serverless GPU training with free credits
• ✅ Unsloth Optimization: 2x faster training, 60% less memory
• ✅ LoRA Fine-tuning: Efficient parameter-efficient training
• ✅ 2000 Training Examples: Real Hide-and-Seek game data
• ✅ Automated Testing: Built-in model evaluation
• ✅ Production Ready: Export merged 16-bit models

📊 Dataset

Source: hider_raw.jsonl (2000 examples)

Format:

{
  "messages": [
    {"role": "system", "content": "You are a HIDER agent..."},
    {"role": "user", "content": "[4.232, 1.478, 0.000, ...]"},
    {"role": "assistant", "content": "[8, 0, 7, 1, 1]"}
  ]
}

💰 Cost

Using Modal's free tier ($30 credits):

• Training: ~$0.40 per run (20-30 mins on T4)
• Testing: ~$0.05 per run (2-3 mins)
• You can train 75+ times with free credits!

📚 Documentation

• SETUP.md - Detailed setup guide
• train_modal.py - Training script with comments
• test_model.py - Testing script with samples

🛠️ Tech Stack

• Model: Gemma 3 270M (Google)
• Cloud: Modal (serverless GPU)
• Optimization: Unsloth (2x speedup)
• Method: LoRA fine-tuning
• Framework: Transformers, TRL, PEFT

🎓 Learn More

• Modal Documentation
• Gemma 3 Model Card
• Unsloth GitHub

---

Made with ❤️ using Modal's free GPU credits