Modelpublic

RichardErkhov/xsanskarx_-_qwen2-0.5b_numina_math-instruct-awq

sourceHugging Faceupdated 2y agoView on Hugging Face

0likes4downloads

Model Card

Quantization made by Richard Erkhov.

Github

Discord

Request more models

qwen2-0.5bnuminamath-instruct - AWQ

—Model creator: https://huggingface.co/xsanskarx/
—Original model: https://huggingface.co/xsanskarx/qwen2-0.5bnuminamath-instruct/

Original model description: --- datasets:

—AI-MO/NuminaMath-CoT language:
—en library_name: transformers license: cc-by-4.0 tags:
—text-generation-inference
—chat
—qwen2
—conversational
—math
—maths
—unsloth
—trl
—sft ---

xsanskarx/qwen2-0.5bnuminamath-instruct

This repository contains a fine-tuned version of the Qwen-2 0.5B model specifically optimized for mathematical instruction understanding and reasoning. It builds upon the Numina dataset, which provides a rich source of mathematical problems and solutions designed to enhance reasoning capabilities even in smaller language models.

Motivation

My primary motivation is the hypothesis that high-quality datasets focused on mathematical reasoning can significantly improve the performance of smaller models on tasks that require logical deduction and problem-solving. Uploading benchmarks is the next step in evaluating this claim.

Model Details

—Base Model: Qwen-2 0.5B
—Fine-tuning Dataset: Numina COT
—Key Improvements: Enhanced ability to parse mathematical instructions, solve problems, and provide step-by-step explanations.

Usage

You can easily load and use this model with the Hugging Face Transformers library:

python

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("xsanskarx/qwen2-0.5b_numina_math-instruct")
model = AutoModelForCausalLM.from_pretrained("xsanskarx/qwen2-0.5b_numina_math-instruct")