fourbic/disarm-ew-llama3-finetuned
The fourbic/disarm-ew-llama3-finetuned model is an 8 billion parameter Llama-3.1 variant, fine-tuned by fourbic using LoRA for specialized analysis of election-related content. Optimized for Apple Silicon, this model excels at classifying disinformation, misinformation, and coordinated influence operations within Nigerian election content. It processes inputs up to 32768 tokens, focusing on DISARM Framework classification and narrative analysis.
Loading preview...
DISARM Election Watch - Fine-tuned Llama-3.1 Model
This model is a specialized 8 billion parameter Llama-3.1 variant, fine-tuned by fourbic to analyze election-related content, particularly from Nigeria. It leverages LoRA (Low-Rank Adaptation) for efficient fine-tuning and is specifically optimized for Apple Silicon (M1 Max) hardware, ensuring fast local inference.
Key Capabilities
- Disinformation Analysis: Identifies and classifies disinformation, misinformation, and coordinated influence operations.
- DISARM Framework: Optimized for classifying content according to DISARM Framework techniques and meta-narratives.
- Narrative Analysis: Extracts key indicators and categorizes content related to undermining electoral institutions.
- Local Deployment: Designed for privacy, speed, and customization through local deployment via MLX-LM or Ollama.
- Hardware Optimization: Benefits from Metal GPU acceleration and efficient memory management on Apple Silicon.
Training and Performance
The model was trained on 6,019 examples of Nigerian election content, achieving a final training loss of 1.064. It demonstrates an inference speed of approximately 20 tokens/second with a memory usage of 16.149 GB. The LoRA adapters are lightweight at 1.7MB, while the fused model is 16GB.
Use Cases
This model is ideal for researchers, journalists, and organizations monitoring election integrity, providing a robust tool for automated content classification and narrative detection in election-related discourse.