Modelpublic

ahmetggg/Luck-Qwen2.5-Coder-3B-STEM-GGUF

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes154downloads

Model Card

Luck-Qwen2.5-Coder-3B-STEM (GGUF Q4KM)

A fine-tuned version of Qwen2.5-Coder-3B-Instruct specialized in STEM and code-related tasks. This model has been trained on a curated mix of code, mathematics, and development-focused datasets with emphasis on Python, Kotlin, Unity (C#), and mathematical reasoning.

⚠️ Important Notice

This is an experimental fine-tune. While the model shows improvements in certain STEM domains, it has known limitations:

—May occasionally enter repetition loops on complex prompts
—Can produce verbose responses with unnecessary explanations
—HTML/JS generation quality is lower than the base model
—Best used with temperature=0.2 and max_new_tokens=512 for stable outputs

For production use, we recommend the original Qwen/Qwen2.5-Coder-3B-Instruct.

Model Details

—Base Model: Qwen2.5-Coder-3B-Instruct
—Training Steps: 1,500
—Quantization: Q4KM (~2.0 GB)
—Format: GGUF
—Training Framework: Unsloth + PEFT (LoRA)
—Target Languages: Python, Kotlin, C# (Unity), JavaScript, SQL

Training Data Mix

Dataset	Samples	Domain
saurabh5/rlvr-code-data-Kotlin	2,500	Kotlin
ise-uiuc/Magicoder-Evol-Instruct-110K	4,000	Code Instruct
theblackcat102/evol-codealpaca-v1	2,500	Code Alpaca
vishnuOI/unity-dev-instructions	2,500	Unity/C#
bigcode/python-stack-v1-functions-filtered-sc2	1,500	Python
ryanmarten/OpenThoughts-1k-sample	500	Reasoning
TIGER-Lab/MathInstruct	1,500	Mathematics

Quick Start with Ollama

1. Download the model

bash

huggingface-cli download ahmetggg/Luck-Qwen2.5-Coder-3B-STEM-GGUF --include "*.gguf" --local-dir ./models