🌌 plexus-accelerator-ner-v3
plexus-accelerator-ner-v3 is a high-performance, multilingual Named Entity Recognition (NER) engine. It serves as the critical Triage and Accelerator Layer for the Plexus V4.2 Adaptive Intelligence Engine, a system designed to assist users with executive dysfunction and ADHD through proactive AI coaching.
🏗 System Role: The "Accelerator" Layer
In the Plexus V4.2 architecture, this model sits at Layer 1 (The Accelerator). Its primary goal is to resolve user intent with minimal computational overhead.
- Semantic Routing: Extracts entities (Project Names, Deadlines, Categories) to create anonymized intent templates.
- LLM Bypass: Allows the system to execute "Automatic Paths" (>98% confidence) without invoking expensive LLMs like GPT-4o, reducing latency by ~85%.
- Privacy-First: Operates as a "Data Sanitizer," ensuring that personally identifiable information (PII) is tokenized before being passed to cloud-based reasoning layers.
📊 Technical Specifications
Model Card
| Feature | Specification |
|---|---|
| Base Model | bert-base-multilingual-cased |
| Parameters | 177M |
| Language Support | Multilingual (English, German, French, Arabic, etc.) |
| Input Window | 512 Tokens |
| Latency | ~40ms on Standard CPU |
Training Hyperparameters
The model was fine-tuned using a high-precision, curated dataset of executive function commands and organizational logic.
- Optimizer: AdamW with Weight Decay (0.01)
- Learning Rate: $2 \times 10^{-5}$
- Scheduler: Linear warmup with decay
- Epochs: 5
- Batch Size: 16 (on NVIDIA T4 GPU)
📈 Performance Metrics (Test Set)
The model demonstrates state-of-the-art accuracy for domain-specific entity extraction in the productivity space.
| Metric | Score | Note |
|---|---|---|
| Accuracy | 0.942 | Global token accuracy |
| Precision | 0.915 | Entity-level precision |
| Recall | 0.928 | Sensitivity to complex entities |
| F1-Score | 0.921 | Primary optimization metric |
🛠 Usage & Implementation
Direct Inference
from transformers import AutoTokenizer, AutoModelForTokenClassification, pipeline
# Initialize the Plexus Accelerator
tokenizer = AutoTokenizer.from_pretrained("nour833/plexus-accelerator-ner-v3")
model = AutoModelForTokenClassification.from_pretrained("nour833/plexus-accelerator-ner-v3")
nlp = pipeline("ner", model=model, tokenizer=tokenizer, aggregation_strategy="simple")
# Sample "Guardian Angel" Scenario
prompt = "Remind me to finalize the investment deck before the 4 PM meeting on Friday."
results = nlp(prompt)
for entity in results:
print(f"Entity: {entity['word']} | Label: {entity['entity_group']} | Score: {entity['score']:.4f}")
Entity Schema
The model is fine-tuned to recognize the following specialized labels:
B-PROJECT: Start of a project or goal name.B-TIME: Temporal constraints and deadlines.B-CATEGORY: Task classification (Work, Social, Health).B-URGENCY: Priority markers in natural language.
🛡 Ethical Considerations & Privacy
This model is built on the principle of Local-First AI. By performing NER locally, Plexus minimizes the surface area for data leaks. This is particularly critical for users tracking sensitive medical or professional data.
🔗 Project Context
This model is part of the Plexus Project, an initiative to democratize executive function support.
- Developer: Nour Kooli
- Project Hub: GitHub Repository