实验结果令人振奋:基于OpenMathInstruct-2训练的8B模型在MATH基准测试上的表现比Llama3.1-8B-Instruct模型提高了15.9%,达到67.8%的准确率,成为10B以下参数量中最强的开源模型之一。而70B模型更是达到了71.9%的准确率,超越了Llama3.1-70B-Instruct 3.9个百分点。 这项研究不仅为AI数学能力的提升提供了宝贵的开源资源,也...
(1)I use vllm’s api ,error for example Question: Melanie is a door-to-door saleswoman. She sold a third of her vacuum cleaners at the green house, 2 more to the red house, and half of what was left at the orange house. If Melanie has 5 vacuum cleaners left, how many did she...
This project fine-tunes the Phi-3-mini-4k-instruct LLM using the ArXiv math dataset via QLoRA, optimizing it for domain-specific tasks like mathematical queries. The fine-tuned model is accessible locally through OpenWebUI, providing efficient and accura
Reproduces the problem - command or script python3 run.py --models hf_qwen2_7b_instruct --datasets math_gen Reproduces the problem - error message As the image1 shows, the tested accuracy of Qwen2-7b-Instructed on MATH is 23.76, which is dramatically lower than the reported score on the...