Denisijcu/vertexgemma

Vertex-Gemma Chess (2B QLoRA)

Vertex-Gemma Chess is a specialized tactical chess model fine-tuned from Gemma 2 (2B-Instruct) by Vertex Coders LLC (Miami, FL). Optimized for fast convergence and lightweight local execution, it competes directly in the Vertex Chess Arena architecture duels.

Model Details

• Developer: Vertex Coders LLC (Denis Sanchez Leyva)
• Base Model: google/gemma-2-2b-it
• Fine-tuning Framework: Unsloth QLoRA (Local workstation: NVIDIA GTX 1660 Ti, 6GB VRAM)
• Dataset: Grandmaster Blend (5,000 elite chess positions and tactical sequences)
• Training Hyperparameters:
  • Epochs: 1
  • Total Steps: 80 steps (optimized for Gemma's high token-density convergence)
  • Effective Batch Size: 8
  • Precision: FP16 / LoRA rank ($r=8$)

Intended Use

Engineered for local deployment via Ollama and integrated with the ChessArbiter 1-ply tactical search engine to ensure robust, rule-compliant chess matches without illegal moves or infinite loops.

How to Get Started

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "your-huggingface-username/vertex-gemma-chess"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float16, device_map="auto")