mahernaija/qwen25-32b-nemotron-finetuned

TEXT GENERATIONConcurrency Cost:2Model Size:32.8BQuant:FP8Ctx Length:32kPublished:Mar 29, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

mahernaija/qwen25-32b-nemotron-finetuned is a 32.8 billion parameter language model, a full fine-tune of Qwen/Qwen2.5-32B by mahernaija. It is specifically optimized for step-by-step reasoning across math, code, and science problems, incorporating `` traces. This model excels at generating detailed reasoning processes, showing significant ROUGE-L improvements in these domains while largely preserving general knowledge benchmarks.

Loading preview...

Model Overview

This model, mahernaija/qwen25-32b-nemotron-finetuned, is a full fine-tune of the 32.5 billion parameter Qwen/Qwen2.5-32B base model. It was trained on the Llama-Nemotron Post-Training Dataset to enhance its reasoning capabilities.

Key Capabilities

  • Step-by-step Reasoning: Produces detailed <think> traces for math, code, and science problems, a feature absent in the base model.
  • Improved Domain Performance: Achieved a 76% overall ROUGE-L improvement on Nemotron evaluation samples, with significant gains in science (+87%) and code (+233%).
  • General Knowledge Preservation: Maintained general knowledge benchmarks (MMLU, HellaSwag, Winogrande) with less than 1.2% change, indicating broad capabilities are retained.
  • Robust Training: Fine-tuned for one epoch on 90K diverse samples (40K math, 40K code, 20K science) using 16 NVIDIA H200 GPUs, resulting in a 70% drop in training loss and good generalization.

Ideal Use Cases

This model is particularly well-suited for applications requiring explicit, step-by-step problem-solving in technical domains. It is an excellent choice for tasks involving:

  • Generating detailed explanations for mathematical proofs.
  • Assisting with code debugging or understanding by showing intermediate thought processes.
  • Providing structured reasoning for scientific questions.