YiPz/llama3-8b-pokerbench-sft

Llama 3 8B - PokerBench SFT

Fine-tuned Llama 3.1 8B Instruct for poker decision-making using LoRA, trained on PokerBench dataset.

Training Details

• Base Model: Meta-Llama-3.1-8B-Instruct
• Training Data: PokerBench (RZ412/PokerBench)
• Method: LoRA fine-tuning (merged)
• Training Steps: 5,000
• Batch Size: 128
• Learning Rate: 1e-6

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("YiPz/llama3-8b-pokerbench-sft")
tokenizer = AutoTokenizer.from_pretrained("YiPz/llama3-8b-pokerbench-sft")

messages = [
    {"role": "system", "content": "You are an expert poker player. Respond with your action in <action></action> tags."},
    {"role": "user", "content": "Your poker scenario..."}
]

inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True)
outputs = model.generate(inputs, max_new_tokens=32, temperature=0.1)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Output Format

Actions are returned in <action></action> tags:

• <action>fold</action>
• <action>call</action>
• <action>check</action>
• <action>raise 15</action>
• <action>bet 10</action>

GGUF Versions

Quantized GGUF versions for llama.cpp/Ollama: YiPz/llama3-8b-pokerbench-sft-gguf

License

Subject to Llama 3 license.