--- language: - en license: apache-2.0 library_name: transformers tags: - math - reasoning - qwen2.5 - gguf - qlora - vertex-coders - edge-ai base_model: Qwen/Qwen2.5-0.5B-Instruct --- # Vertex-Math-0.5B (Q8_0) **Vertex-Math-0.5B** is a compact mathematical reasoning language model fine-tuned from `Qwen2.5-0.5B-Instruct` using QLoRA. Optimized specifically for edge devices, offline mobile assistants, and resource-constrained environments, this model delivers rigorous step-by-step mathematical problem-solving while maintaining an ultra-lightweight footprint. ## Model Details - **Base Model:** [Qwen/Qwen2.5-0.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct) - **Architecture:** Qwen2 (0.5B parameters) - **Training Dataset:** [HuggingFaceH4/MATH-500](https://huggingface.co/datasets/HuggingFaceH4/MATH-500) (Filtered and token-length audited) - **Quantization:** Q8_0 (GGUF) - **Developer:** Denis Sanchez Leyva (Vertex Coders LLC) ## Training Configuration & Hyperparameters - **Method:** QLoRA (Rank `r=8`, Alpha `16`) - **Target Modules:** `q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj` - **Hardware:** Local training on NVIDIA GTX 1650 (4GB VRAM) - **Max Sequence Length:** 1024 tokens - **Optimizer:** Paged AdamW 8-bit - **Loss Function:** Fused Cross-Entropy Loss - **Epochs:** 2 (80 steps) ## Prompt Format The model is trained using the ChatML template with a specialized system prompt enforcing step-by-step deductive reasoning: ```text <|im_start|>system You are a compact mathematical reasoning engine optimized for mobile edge devices. Solve problems step by step with precision and conclude with '#### '.<|im_end|> <|im_start|>user [Your math problem here]<|im_end|> <|im_start|>assistant