Walkowicz/kairos-pi-large-1.0

Kairos PI Large 1.0 (~395M)

The Central Disambiguation & Fusion Engine with Native 8,192 Context & 4D Risk Calibration

Kairos PI Large 1.0 is the Tier 2 central reasoning and multimodal fusion engine of the Kairos PI dual-tier architecture. Powered by ModernBERT-large, it consumes long temporal series (up to 8,192 native tokens) in a single forward pass, solving plant-wide SCADA alarm storms and central ADAS domain arbitration without autoregressive decoding delay.

It evaluates 20 to 50+ candidate root causes with cross-option self-attention while simultaneously outputting a calibrated 4-dimensional continuous risk vector (0.0015 MSE).

---

📦 Direct Download

File	Size	Format	Direct Download	Description
model.safetensors	1.60 GB	Safetensors	Download model.safetensors	Official fine-tuned production weights

---

📊 80-Case Cyber-Physical Gauntlet Leaderboard

The complete 80-case evaluation suite is documented in detail in BENCHMARK.md.

Dimension / Metric	🥇 Kairos PI Symbiotic (Dual-Tier)	🥈 Kairos PI Large 1.0 (This Model)	🥉 Kairos PI Low 1.0	4️⃣ Convai Laya Baseline
Overall 80-Case Gauntlet Score	97.5% (78/80)	95.0% (76/80)	92.5% (74/80)	48.8% (39/80)
Active Abstention Recall (20)	100.0% (20/20)	100.0% (20/20)	100.0% (20/20)	0.0% (0/20) (Lacks OOD gate)
Industrial SCADA Wind Power (20)	100.0% (20/20)	100.0% (20/20)	80.0% (16/20)	55.0% (11/20)
Robotics & Flight Telemetry (10)	100.0% (10/10)	100.0% (10/10)	80.0% (8/10)	70.0% (7/10)
Automotive CAN-Bus Telemetry (20)	100.0% (20/20)	85.0% (17/20)	100.0% (20/20)	65.0% (13/20)
IoT Smart Systems Telemetry (10)	100.0% (10/10)	90.0% (9/10)	100.0% (10/10)	80.0% (8/10)
Single-Pass Latency P50 (GPU)	35.8 ms / 168.3 ms	168.3 ms (20 options)	35.8 ms	36.8 ms
Max Context Window	8,192 tokens	8,192 native tokens	2,048 tokens	512 tokens (capped)
Continuous Risk Calibration	4D Vector (0.0015 MSE)	4D Continuous Vector (0.0015 MSE)	1D Calibrated Brier	0D (Discrete only)
Memory Footprint (FP16 / INT8)	299 MB + 799 MB	~799 MB / ~198 MB	~299 MB / ~74 MB	~640 MB / ~160 MB
Target Deployment	Full Industrial Stack	Central ADAS / SCADA Plants	Edge ECU / SoC (<15W)	Cloud Server Only

---

Authentic Cyber-Physical Training Datasets

• Vestas V52 Industrial Wind Turbine SCADA (Granada from vossmoos/vestasv52-scada-windturbine-granada): 40,000 real records across 23 channels (WindSpeed, Power, Pitch, Rotor/Generator RPM, Bearing Temperatures, Gear Oil Temperatures, Relative Yaw Error).
• Avionics & Aerial Robotics (npeng/drone-telemetry): 190-channel flight series capturing antenna attenuation, strong wind warnings, and 3D kinematics.
• Physical Conservation Invariant Violations: Deliberate physical anomalies (rotor at 2,800 RPM with 0.0 m/s wind) for 100% active abstention recall.

---

How to Download & Run

Method 1: Hugging Face CLI Download

huggingface-cli download Walkowicz/kairos-pi-large-1.0 --local-dir ./kairos-pi-large-1.0

Method 2: Git Clone

git clone https://huggingface.co/Walkowicz/kairos-pi-large-1.0

Method 3: Python Inference with 4D Risk Calibration

import torch
from transformers import AutoTokenizer
from huggingface_hub import hf_hub_download
import sys
from pathlib import Path

repo_id = "Walkowicz/kairos-pi-large-1.0"
tokenizer = AutoTokenizer.from_pretrained(repo_id)

model_file = hf_hub_download(repo_id=repo_id, filename="modeling_kairos_pi_large.py")
sys.path.append(str(Path(model_file).parent))
from modeling_kairos_pi_large import KairosPILargeForDecision

model = KairosPILargeForDecision.from_pretrained(repo_id).to("cuda" if torch.cuda.is_available() else "cpu")
model.eval()

candidate_options = [
    "NOMINAL_GRID_POWER_GENERATION",
    "GEARBOX_BEARING_COOLING_AND_TORQUE_DERATING",
    "YAW_SYSTEM_REALIGNMENT_AND_INSPECTION",
    "AERODYNAMIC_PITCH_FEATHERING_EMERGENCY_STOP",
    "STATUS_ABSTAIN"
]

telemetry = (
    "T:14:15:00 WS:8.5m/s PWR:450.0kW PITCH:0.2deg GEN:1540rpm ROT:18.1rpm GEAR_BEAR:79.8C OIL:71.2C YAW_ERR:0.5deg\n"
    "T:14:15:10 WS:8.4m/s PWR:448.0kW PITCH:0.2deg GEN:1538rpm ROT:18.0rpm GEAR_BEAR:83.4C OIL:73.5C YAW_ERR:0.6deg"
)

prompt = (
    "[CLS] [MULTICHOICE] IEC 61400-25 Wind Turbine Supervisory Control [SEP] "
    + " ".join([f"[MASK] {opt}" for opt in candidate_options])
    + f" [SEP] {telemetry}"
)

inputs = tokenizer(prompt, return_tensors="pt", max_length=1024, truncation=True).to(model.device)
with torch.no_grad():
    decision = model(inputs["input_ids"], inputs["attention_mask"], candidate_names=candidate_options, task="multichoice")

print(f"Status:          {decision.status.upper()}")
print(f"Selected Action: {decision.selected_option}")
print(f"Confidence:      {decision.confidence * 100:.2f}%")
print(f"Latency:         {decision.latency_ms:.2f} ms")
print(f"Risk Vector:     Severity={decision.risk_vector['severity_index']:.2f} | FailProb={decision.risk_vector['failure_probability']:.2f} | RUL={decision.risk_vector['remaining_useful_life_idx']:.2f}")

---

🛡️ License

Apache 2.0. Open for commercial and industrial use.