fourbic/disarm-ew-llama3-finetuned

DISARM Election Watch - Fine-tuned Llama-3.1 Model

Model Description

This is a fine-tuned version of the Llama-3.1 model specifically optimized for DISARM Framework analysis of election-related content. The model has been trained on a comprehensive dataset of Nigerian election content from multiple platforms to identify and classify disinformation, misinformation, and coordinated influence operations.

Model Details

• Base Model: ArapCheruiyot/disarm_ew-llama3
• Fine-tuning Method: LoRA (Low-Rank Adaptation)
• Optimization: Apple Silicon (M1 Max) optimized
• Training Data: 6,019 examples from multiple sources
• Task: DISARM Framework classification and narrative analysis
• Language: English
• License: MIT

Training Configuration

• LoRA Rank: 16
• Batch Size: 1
• Learning Rate: 3e-4
• Sequence Length: 2048
• Training Iterations: 600
• Final Training Loss: 1.064
• Final Validation Loss: 1.354
• Framework: MLX-LM
• Hardware: Apple M1 Max (64GB RAM)

Quick Start

Using with MLX-LM

from mlx_lm import load, generate

# Load the complete fine-tuned model
model, tokenizer = load("models/disarm_ew_llama3_finetuned")

# Example prompt
prompt = """### Instruction:
Classify the following content according to DISARM Framework techniques and meta-narratives:

### Input:
A viral WhatsApp broadcast claims that the BVAS machines have been pre-loaded with votes by INEC in favour of the incumbent party.

### Response:"""

# Generate response
response = generate(model, tokenizer, prompt, max_tokens=256, temp=0.1)
print(response)

Using with Ollama

# Create Ollama model
ollama create disarm-ew-llama3-finetuned -f Modelfile

# Run the model
ollama run disarm-ew-llama3-finetuned "Your prompt here"

Example Usage

ollama run disarm-ew-llama3-finetuned "### Instruction:
Classify the following content according to DISARM Framework techniques and meta-narratives:

### Input:
A viral WhatsApp broadcast claims that the BVAS machines have been pre-loaded with votes by INEC in favour of the incumbent party.

### Response:"

Expected Output

{
  "meta_narrative": "Compromised Election Technology",
  "primary_disarm_technique": "T0022.001: Develop False Conspiracy Theory Narratives about Electoral Manipulation and Compromise",
  "confidence_score": 0.98,
  "key_indicators": ["BVAS", "pre-loaded", "INEC"],
  "platform": "WhatsApp",
  "language": "en",
  "category": "Undermining Electoral Institutions"
}

Performance

Training Performance

• Training Loss: 1.064
• Validation Loss: 1.354
• Training Speed: ~1.16 iterations/second
• Memory Usage: 19.161 GB peak during training

Inference Performance

• Inference Speed: ~20 tokens/second
• Memory Usage: 16.149 GB during inference
• Model Size: 16GB (fused), 1.7MB (LoRA adapters)

Hardware Optimization

• Apple Silicon: Optimized for M1 Max
• Metal GPU: Accelerated inference
• Memory Management: 16GB wired memory optimization

Model Files

Fused Model (Complete)

• Size: 16GB
• Format: MLX-LM safetensors
• Files: 4 model weight files + configuration

LoRA Adapters (Lightweight)

• Size: 1.7MB
• Format: safetensors
• Files: Final adapters + training checkpoints

Local Deployment Benefits

• Privacy: Run locally without sending data to external servers
• Speed: Fast inference on local hardware
• Customization: Modify prompts and parameters as needed
• Offline: Works without internet connection

Contact

For questions, issues, or collaboration opportunities:

• Model Repository: ArapCheruiyot/disarm-ew-llama3-finetuned
• Dataset Repository: ArapCheruiyot/disarm-election-watch-dataset