DuoNeural/Qwen2.5-Coder-3B-SFT-WebCode

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:3.1BQuant:BF16Ctx Length:32kPublished:May 4, 2026License:apache-2.0Architecture:Transformer Open Weights Warm

DuoNeural/Qwen2.5-Coder-3B-SFT-WebCode is a 3.1 billion parameter SFT fine-tune by DuoNeural, based on the Qwen2.5-Coder-3B-Instruct model. This model is specifically fine-tuned using the DuoNeural/Gemma4-E2B-SFT-WebCode dataset, focusing on code-related tasks. It is designed for exploring emergent behaviors in small language models, particularly in code generation and understanding contexts.

Loading preview...

Overview

DuoNeural/Qwen2.5-Coder-3B-SFT-WebCode is a 3.1 billion parameter language model, fine-tuned by DuoNeural. It is built upon the Qwen/Qwen2.5-Coder-3B-Instruct base model and has undergone Supervised Fine-Tuning (SFT) using the specialized DuoNeural/Gemma4-E2B-SFT-WebCode dataset. This model is part of DuoNeural's research into emergent behaviors within small language models.

Training Details

The model was trained using LoRA with a rank of 16 and an alpha of 32, over 3 epochs, with a learning rate of 2e-4 and an effective batch size of 16.

Benchmark Results

Evaluation was conducted using lm_eval 0.4.x on GSM8K and ARC-Challenge datasets. While the model shows a decrease in GSM8K performance compared to its baseline, its ARC-Challenge scores remain consistent.

Model GSM8K flex ARC-norm ARC-acc
Baseline 0.5807 0.4957 0.4590
Qwen2.5-Coder-3B-SFT-WebCode 0.3207 0.4957 0.4590

Use Cases

This model is suitable for developers and researchers interested in:

  • Code-related tasks: Leveraging its SFT on a web code dataset.
  • Exploring small language model capabilities: Particularly in the context of code generation and understanding.
  • Further fine-tuning: As a base for more specialized code-centric applications.