FrederickSundeep/nova2-14b

Nova2-14B: A Persona-Driven AI Assistant

Nova2-14B is a 14.7 billion parameter large language model developed by Frederick Sundeep Mallela. It is fine-tuned from the powerful Qwen/Qwen3-14B base model using QLoRA and Unsloth, resulting in a fully merged, standalone model that requires no adapter dependencies for inference. While the base model supports up to 40K tokens, Nova2-14B is fine-tuned with a maximum sequence length of 2048 tokens.

Key Differentiators & Capabilities

Nova2-14B maintains all the robust capabilities of its Qwen3-14B base, including:

• Code generation: Supports Python, JavaScript, C++, SQL, and more.
• Reasoning and Math: Excels in logical problem-solving and advanced mathematics.
• Multilingual support: Inherits support for over 100 languages.
• Consistent Persona: Responds as "Nova," an AI assistant created by Frederick, with a consistent identity and tone.
• Instruction following: Designed for precise task execution and fully supports custom system prompts.
• Tool use: Compatible with function calling.

Training and Optimization

The model was fine-tuned using a custom dataset focused on establishing Nova's identity, technical knowledge, and personality. This process involved Supervised Fine-Tuning (SFT) with QLoRA, utilizing a Tesla T4 GPU. The fine-tuning specifically trained Nova to never reveal its underlying architecture details, making it suitable for integrated application use like the NovaMind chat application.

Intended Use Cases

• Powering the NovaMind AI chat application.
• General-purpose AI assistant tasks.
• Code generation and debugging.
• Technical question answering.
• As a base model for further fine-tuning.

Limitations

• Not evaluated on standard benchmarks post fine-tuning.
• May occasionally revert to base Qwen3 behavior due to a relatively small custom dataset.
• Context limited to 2048 tokens in its fine-tuned configuration, despite the base model's 40K capability.