Belaleatsbanana/qwen25-coder-32b-sft-ocr2-combined

Hugging Face
TEXT GENERATIONConcurrency Cost:2Model Size:32.8BQuant:FP8Ctx Length:32kPublished:May 17, 2026License:apache-2.0Architecture:Transformer Open Weights Warm

Belaleatsbanana/qwen25-coder-32b-sft-ocr2-combined is a 32.8 billion parameter language model fine-tuned from Qwen/Qwen2.5-Coder-32B-Instruct. It specializes in competitive programming, generating detailed chain-of-thought reasoning traces and corresponding Python or C++ solutions. The model was trained using supervised fine-tuning on the nvidia/OpenCodeReasoning-2 dataset, making it highly effective for complex coding challenges with a 32,768 token context length.

Loading preview...

Model Overview

Belaleatsbanana/qwen25-coder-32b-sft-ocr2-combined is a powerful 32.8 billion parameter language model, specifically fine-tuned for competitive programming tasks. It builds upon the robust Qwen2.5-Coder-32B-Instruct base model, enhancing its capabilities through supervised fine-tuning (SFT).

Key Capabilities

  • Advanced Code Reasoning: Generates comprehensive chain-of-thought reasoning traces for competitive programming problems.
  • Multi-language Solution Generation: Produces high-quality Python or C++ code solutions based on the problem description and reasoning.
  • Extensive Context Window: Supports a maximum sequence length of 32,768 tokens, allowing for processing of complex problems and long reasoning paths.
  • Specialized Training: Fine-tuned on the nvidia/OpenCodeReasoning-2 dataset, which combines various competitive programming sources like TACO, APPS, CodeContests, and Codeforces.

Training Details

The model was trained using QLoRA (4-bit NF4, r=64, alpha=128) and then merged back to full precision. This method efficiently leverages the extensive and verified reasoning traces from the OpenCodeReasoning-2 dataset, which includes problems with solutions verified by code execution.

Ideal Use Cases

This model is particularly well-suited for:

  • Competitive Programming Assistance: Generating solutions and detailed thought processes for coding challenges.
  • Code Education: Providing step-by-step reasoning for complex algorithms and data structures.
  • Automated Code Generation: Creating functional code snippets in Python or C++ from problem statements.

Finetunes
1 models