TAKTKRONE-I: Metro Operations Control Center Language Model

Model Description

TAKTKRONE-I is a specialized language model fine-tuned for metro operations control center (OCC) assistance. Based on Llama 3.1 8B Instruct, it has been trained on realistic operational scenarios to help transit operators with incident response, diagnosis, and recovery planning.

Model Details

• Model Type: Instruction-tuned Llama 3.1 8B with LoRA adapters
• Base Model: meta-llama/Llama-3.1-8B-Instruct
• Training Method: Supervised Fine-Tuning (SFT) with LoRA
• Training Data: 5,247 real OCC dialogue samples
• Languages: English
• License: Apache 2.0
• Developer: Gustav Olaf Yunus Laitinen-Fredriksson Lundström-Imanov
• Version: 1.0
• Model Size: 8.03B parameters (LoRA adapters: 67.1M)
• DOI: 10.57967/hf/8167

Performance

Overall Metrics

• Overall Quality Score: 0.829/1.0 (Grade: B+)
• API Latency: P95 1,247ms, P50 456ms
• Safety Compliance: 98.7%
• System Uptime: 99.87%

Benchmark Results

OCC Summarization

• ROUGE-L: 0.823
• BLEU: 0.687
• Factual Consistency: 0.942
• Readability Score: 0.889

Disruption Diagnosis

• Diagnosis Accuracy: 0.891
• Recommendation Relevance: 0.847
• Confidence Calibration: 0.785
• Response Time: 1.247s (P95)

Safety & Compliance

• Safety Violation Rate: 0.008%
• Guard Effectiveness: 0.987
• False Positive Rate: 0.042

Training Data

The model was trained on operational scenarios including:

• Signal failures: 1,234 cases across MTA, MBTA, WMATA
• Power outages: 567 cases with emergency protocols
• Medical emergencies: 891 cases with EMS coordination
• Track obstructions: 423 cases with maintenance responses
• Weather impacts: 678 cases with service adjustments
• Security incidents: 234 cases with law enforcement
• Equipment failures: 1,260 cases with various systems

Intended Use

Primary Use Cases

• Incident summarization and reporting
• Disruption cause analysis and diagnosis
• Recovery strategy recommendation ranking
• Real-time decision support for OCC operators
• Training scenario generation for staff education

Out-of-Scope Uses

• Autonomous control of transit systems
• Financial or business decision making
• Medical diagnosis or emergency response
• Legal advice or compliance decisions

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# Load model and tokenizer
model_name = "olaflaitinen/taktkrone-i"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.float16,
    device_map="auto"
)

# Example query
query = "Signal failure reported at Times Square, Line 4/5/6. Multiple trains held. What's the recommended response?"

# Tokenize and generate
inputs = tokenizer(query, return_tensors="pt")
outputs = model.generate(
    **inputs,
    max_new_tokens=256,
    temperature=0.7,
    do_sample=True
)

response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Model Architecture

• Base Architecture: Llama 3.1 8B Instruct
• Fine-tuning: LoRA (Low-Rank Adaptation)
• LoRA Rank: 16
• LoRA Alpha: 32
• Target Modules: Query and Value projection layers
• Adapter Parameters: 67.1M (0.84% of base model)

Deployment Requirements

Hardware

• Minimum: 16GB GPU memory, 32GB RAM
• Recommended: 40GB GPU memory, 64GB RAM
• Inference: NVIDIA A100/H100 or equivalent

Software Dependencies

• PyTorch >= 2.1.0
• Transformers >= 4.36.0
• vLLM >= 0.2.7 (for production serving)
• CUDA >= 11.8

Limitations

• Trained primarily on subway/metro systems (limited bus/ferry scenarios)
• English language only
• Performance may vary on novel incident types not seen in training
• Requires human oversight for all critical operational decisions
• Not suitable for real-time control system integration

Ethical Considerations

Safety Measures

• Multi-layer safety guardrails prevent harmful outputs
• Confidence scoring indicates when human review is required
• Built-in bias detection for fair treatment across demographics
• Audit logging for all operational recommendations

Data Privacy

• All training data anonymized and aggregated
• No personally identifiable information (PII) retained
• Compliance with transportation data protection regulations

Evaluation

The model has been evaluated on 6 comprehensive benchmarks:

1. OCC Summarization - Incident report summarization
2. Disruption Diagnosis - Root cause analysis
3. Recovery Ranking - Action prioritization
4. Topology Consistency - Network knowledge validation
5. Safety Guards - Harmful content detection
6. Retrieval QA - Knowledge retrieval accuracy

Citation

@misc{gustav_olaf_yunus_laitinen-fredriksson_imanov_2026,
    author       = { Gustav Olaf Yunus Laitinen-Fredriksson Imanov },
    title        = { taktkrone-i (Revision efdde52) },
    year         = 2026,
    url          = { https://huggingface.co/olaflaitinen/taktkrone-i },
    doi          = { 10.57967/hf/8167 },
    publisher    = { Hugging Face }
}

Contact

• Developer: Gustav Olaf Yunus Laitinen-Fredriksson Lundström-Imanov
• Email: dev@taktkrone.ai
• Website: https://taktkrone.ai
• Repository: https://github.com/olaflaitinen/taktkrone-i

DOI

10.57967/hf/8167

License

Apache 2.0