nmalinowski/pauper-llama3-8b

Pauper Llama 3 8B

Fine-tuned version of meta-llama/Meta-Llama-3-8B-Instruct specialized for Magic: The Gathering's Pauper format using LoRA fine-tuning.

📦 Available Formats

This repository contains both the full HuggingFace model and GGUF quantizations for various use cases.

HuggingFace Transformers (Full Precision)

Perfect for:

• Further fine-tuning
• Maximum quality inference
• Integration with transformers library

GGUF Quantized Models (llama.cpp compatible)

Perfect for:

• LM Studio, Ollama, llama.cpp
• Local inference on consumer hardware
• Faster inference with minimal quality loss

File	Size	Description	Best For
gguf/pauper_llama3_q4km.gguf	~5GB	4-bit quantized	Recommended - Best balance
gguf/pauper_llama3_q5km.gguf	~6GB	5-bit quantized	Better quality
gguf/pauper_llama3_q8.gguf	~8GB	8-bit quantized	Near-original quality
gguf/pauper_llama3_fp16.gguf	~15GB	Full precision	Maximum quality

🚀 Usage

Option 1: HuggingFace Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "nmalinowski/pauper-llama3-8b",
    torch_dtype=torch.float16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("nmalinowski/pauper-llama3-8b")

prompt = "What are the best cards in Pauper?"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Option 2: LM Studio (GGUF - Easiest!)

1. Download gguf/pauper_llama3_q4km.gguf from Files tab
2. Open LM Studio → Load Model
3. Select the downloaded GGUF file
4. Start chatting about Pauper!

Option 3: llama.cpp

# Download the quantized model
huggingface-cli download nmalinowski/pauper-llama3-8b gguf/pauper_llama3_q4km.gguf --local-dir ./

# Run inference
./llama-cli -m pauper_llama3_q4km.gguf \
    -p "What are the top Pauper decks in the current meta?" \
    -n 256 \
    --temp 0.7

Option 4: Ollama

# Create Modelfile
cat > Modelfile <<EOF
FROM ./gguf/pauper_llama3_q4km.gguf
PARAMETER temperature 0.7
PARAMETER top_p 0.9
SYSTEM "You are an expert on Magic: The Gathering's Pauper format."
EOF

# Create and run
ollama create pauper-llama3 -f Modelfile
ollama run pauper-llama3 "Explain the current Pauper meta"

🎯 Training Details

• Base Model: Llama 3 8B Instruct
• Training Method: LoRA (Low-Rank Adaptation)
• Domain: Magic: The Gathering - Pauper format
• LoRA Configuration:
  • Rank: 16
  • Alpha: 32
  • Target modules: q_proj, v_proj
  • Dropout: 0.05

💡 Recommendations

• For most users: Download gguf/pauper_llama3_q4km.gguf and use with LM Studio
• For best quality: Use the full HuggingFace model with transformers
• For low VRAM: Use Q4_K_M quantization (~5GB)
• For high VRAM: Use Q8_0 or FP16 for better quality

📊 Performance

The Q4_K_M quantization offers:

• ✅ ~95% of full precision quality
• ✅ 70% smaller file size
• ✅ Faster inference on CPU and GPU
• ✅ Runs on consumer hardware (16GB RAM recommended)

🎮 Example Prompts

"What are the best removal spells in Pauper?"
"Build me a Pauper deck around Monastery Swiftspear"
"Explain the differences between Affinity and Elves in Pauper"
"What are the current tier 1 Pauper decks?"

⚠️ Limitations

• Specialized for Pauper format - may not perform well on other MTG formats
• May occasionally hallucinate card names or abilities
• Knowledge cutoff: January 2025
• Not suitable for medical, legal, or financial advice

📄 License

This model inherits the Llama 3 Community License from Meta. See LICENSE for details.

🙏 Acknowledgments

• Base model: Meta's Llama 3 8B Instruct
• Training framework: HuggingFace Transformers + PEFT
• Quantization: llama.cpp

📞 Issues & Feedback

If you encounter issues or have suggestions, please open an issue on the Community tab.