Mohamed132411/Qwen3-4B-FitGPT-AR-EN-Instruct

TEXT GENERATIONConcurrency Cost:1Model Size:4BQuant:BF16Ctx Length:32kTool Calling:SupportedPublished:Apr 16, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

Mohamed132411/Qwen3-4B-FitGPT-AR-EN-Instruct is a 4 billion parameter bilingual fitness AI model developed by Mohamed Ramadan. Fine-tuned on a custom-merged Qwen3-4B base, it provides science-based fitness and nutrition guidance in both Arabic and English. This model excels at strict instruction-following, including generating structured JSON output for agent-based systems, making it suitable for precise, automated applications in the fitness domain.

Loading preview...

Overview

Mohamed132411/Qwen3-4B-FitGPT-AR-EN-Instruct is a 4 billion parameter specialized bilingual fitness AI model, developed by Mohamed Ramadan. It is built upon a custom DARE-TIES merge of Qwen3-4B-Instruct-2507 and Qwen3-4B base models, then fine-tuned using Unsloth and Hugging Face TRL. The model is provided as full 16-bit merged weights, ready for direct deployment.

Key Capabilities

  • Bilingual Fitness Expert: Delivers detailed, science-backed advice on workout programming, nutrition planning, exercise technique, and recovery in both Arabic and English.
  • Strict Agent / JSON Mode: Trained to follow formatting instructions precisely, returning only valid JSON, numbers, or lists without unsolicited commentary or markdown wrappers.
  • Arabic-Native Support: Fine-tuned with a dedicated Arabic fitness corpus (CIDAR, alpaca-gpt4-arabic, and custom data) to provide fluent and natural responses in Arabic, unlike many models where Arabic is an afterthought.

Training Details

The model underwent a 3-stage pipeline: foundation merging using DARE-TIES, supervised fine-tuning with ~7,000 curriculum-ordered samples (including diverse English and Arabic fitness datasets), and final weight integration. Datasets included chibbss/fitness-chat, onurSakar/GYM-Exercise, its-myrto/fitness-QA, Varick/workout-routine, hammam/fitness-qa, arbml/CIDAR, alpaca-gpt4-arabic, mlabonne/FineTome-100k, and custom agent examples.

Good For

  • Applications requiring precise, science-based fitness and nutrition advice in English or Arabic.
  • Agent-based systems that need strict instruction following and reliable JSON output.
  • Developers looking for a deployment-ready model without additional adapter loading.

Finetunes
2 models