nooruiit-864/qwen2.5-1.5b-base-ai-safety-domain-lora
The nooruiit-864/qwen2.5-1.5b-base-ai-safety-domain-lora model is a 1.5 billion parameter Qwen2.5-based causal language model, fine-tuned via QLoRA on an AI Safety / Deepfake Misinformation corpus. This merged variant is optimized for memory-efficient deployment, significantly reducing GPU memory footprint while retaining domain-adapted knowledge. It is intended for educational and research use in studying LLM behavior on AI safety and misinformation-related text generation.
Loading preview...
Model Overview
This model, developed by nooruiit-864, is a 1.5 billion parameter Qwen2.5-based causal language model that has undergone domain adaptation. It was fine-tuned using QLoRA on a specialized AI Safety / Deepfake Misinformation corpus and subsequently merged for standalone deployment, eliminating the need for a separate adapter during inference.
Key Features and Optimizations
- Domain-Adapted Knowledge: Specifically trained on AI Safety and Deepfake Misinformation content, enhancing its relevance for these domains.
- Memory Efficiency: The merged and 4-bit NF4 quantized variant significantly reduces peak GPU memory usage by approximately 62% compared to the base model, making it suitable for memory-constrained environments.
- Standalone Deployment: The LoRA adapter is merged into the base weights, allowing for direct loading and inference without additional components.
Performance Highlights (4-bit NF4 Quantized)
- Reduced Memory Footprint: Achieves a peak GPU memory usage of 1.164 GB, down from 3.103 GB for the base model.
- Balanced Inference Speed: Maintains a reasonable inference speed of 14.88 tokens/sec on a Colab T4 GPU.
Intended Use Cases
This model is primarily designed for educational and research purposes to study the behavior of domain-adapted LLMs, particularly in generating and analyzing text related to AI safety and misinformation. Users should be aware of potential mild catastrophic forgetting on general-purpose tasks due to domain adaptation.