Q25 1.5 B Veo Lu

Introduction

Q25-1.5B-VeoLu is a compact, versatile creative model derived from specialized fine-tuning on the Qwen 2.5-1.5B-Instruct model. It is designed to handle a variety of tasks, including writing, roleplaying, adventuring, and reasoning, while maintaining coherence in its outputs.

Architecture

VeoLu merges several components, each focused on specific tasks:

• Bard (Writing): Utilizes the Fujin dataset for writing.
• Scribe (Roleplaying): Employs the Creative Writing Multiturn dataset.
• Cartographer (Adventuring): Based on the SpringDragon dataset.
• Alchemist (Science/Reasoning): Uses ScienceQA, MedquadQA, and Orca Math Word Problems datasets.

The model has 1.5 billion parameters and operates with a configuration that combines different models using task arithmetic, with weights evenly distributed across layers.

Training

The model was trained on a diverse range of datasets to cover various domains:

• Writing: Fujin (Cleaned/extended Rosier).
• Roleplay: Creative Writing Multiturn.
• Adventuring: SpringDragon.
• Science/Reasoning: ScienceQA, MedquadQA, Orca Math Word Problems.

These datasets provide a balanced foundation, enabling the model to perform creative and reasoning tasks effectively.

Guide: Running Locally

1. Install vLLM:

   pip install vllm
   

2. Serve the Model:
   Run the following command to serve the model:

   vllm serve Alfitaria/Q25-1.5B-VeoLu --max-model-len 16384 --max-num-seqs 1
   

3. Hardware Requirements:
   The full-precision model is 3.5GB. It is recommended to use a cloud service with GPU support, such as Google Cloud Platform, AWS, or Azure, to efficiently run the model.

License

The model is made available by inflatebot with contributions from Allura and Auri. Ensure to review and comply with any specific licensing terms provided by the creators when using the model.