🏛️ Millennium-Jev-1.5B: Flagship Non-Autoregressive System 1 Decision Model

Millennium-Jev-1.5B (千禧年·Jev 旗舰版) is the flagship decision reflex model of the Millennium Open-Source Ecosystem (formerly code-named Aegis-S1). Built upon Qwen2.5-1.5B-Instruct, it features 1.54B backbone parameters coupled with 77.08M trainable decision parameters (36.93M LoRA adapter + 40.15M non-autoregressive Cross-Attention decision head).

With Vectorized Option Span Mean-Pooling, Native C++ SDPA Bidirectional Attention, and Split-Conformal Prediction ($\alpha=0.05$), Millennium-Jev-1.5B delivers sub-50ms deterministic decisions with tighter uncertainty bounds, higher autonomous execution rates, and rigorous mathematical safety guarantees.

🌟 Key Highlights

• Enhanced Autonomous Act Rate: Achieves 74.54% autonomous decision pass rate on state-disjoint test sets (+2.40% over the 0.5B baseline) under strict $1-\alpha=95%$ safety bounds.
• Lower Selective Risk: Selective risk on autonomous action dropped to 5.52% (an 11% relative reduction in errors compared to 0.5B's 6.22%).
• Sub-50ms Decision Reflex: Evaluates dynamic runtime candidate options in a single non-autoregressive forward pass, eliminating iterative autoregressive token generation latency.
• Single-Forward Multi-Query Evaluation: Solves multiple heterogeneous typed queries (Choice, Score, Noul) simultaneously in a single prompt.
• Finite-Sample Mathematical Safety (Split-Conformal Prediction): Provides exchangeable marginal coverage $\mathbb{P}(Y \in \mathcal{C}(X)) \ge 94.09%$ with average prediction set size of 1.366.
• Cryptographic Provenance Verification: Bound to SHA-256 checkpoint verification (458c3b73...) preventing silent fallback to uncalibrated legacy weights.

---

📊 Official Benchmark Results (Disjoint Held-Out Set, N=8,657)

Evaluated under strict state-disjoint conditions (zero template or state overlap between train, calibration, and test splits):

Metric	Millennium-Jev 1.5B (Flagship)	Millennium-Jev 0.5B (V6)	Baseline (Raw Qwen Zero-Shot)
Top-1 Generalization Accuracy	84.88%	84.28%	68.60%
Conformal Coverage ($\alpha=0.05$)	94.09%	94.16%	N/A (uncalibrated)
Tri-Gate Autonomous Act Rate	74.54%	72.14%	N/A (no gate)
Selective Risk on Act	5.52%	6.22%	N/A
Abstention / Escalation Rate	25.46%	27.86%	N/A
Average Prediction Set Size	1.366	1.390	N/A

Vertical Domain Performance Breakdown

• Agent Tool Routing (REST/SQL/Shell): 100.00% Top-1 Accuracy, 100.00% Act Rate (0 errors).
• High-Cardinality Actions (20+ candidates): 100.00% Top-1 Accuracy, 99.93% Act Rate.
• Financial Banking Intent (Banking77): 97.38% Top-1 Accuracy, 96.38% Autonomous Act Rate.
• Chinese Broad Domain NLU (TNEWS 15-class): 68.03% Top-1 Accuracy (+2.79% over 0.5B), 59.92% Act Rate (+7.91%).
• Adversarial & Guardrail Safety (BeaverTails): 68.71% Top-1 Accuracy, 28.70% Act Rate (+9.39%), safely escalating ambiguous/adversarial attacks.

---

⚡ Quickstart

Installation

pip install open-s1

Python Inference with 1.5B Flagship

import open_s1 as s1
from open_s1.primitives import Choice, Score, Noul

# Load 1.5B Flagship model
router = s1.load(
    model_dir="sunziqin/millennium-jev-1.5b",
    base_model_path="Qwen/Qwen2.5-1.5B-Instruct",
)

# 1. Single Decision Routing with Conformal Risk Tri-Gate
result = router.decide(
    state="System alert: Disk IO latency spiked to 450ms on database node replica-03.",
    question="Select automated remediation action:",
    candidates=[
        "Restart entire database cluster",
        "Drain node replica-03 and failover to standby",
        "Ignore alert as transient noise",
    ],
    alpha=0.05,
)

print("Selected:", result["selected_option"])
print("Can Act:", result["can_act"])            # True
print("Confidence:", result["confidence"])      # e.g., 0.984
print("Prediction Set:", result["conformal_set"])

# 2. Parallel Multi-Query Forward (Single Forward Pass)
schema = {
    "intent": Choice(["order_query", "refund", "complaint"], question="Primary customer intent?"),
    "urgency": Score(min_val=1.0, max_val=5.0, steps=5, question="Urgency score"),
    "legal_threat": Noul(question="Does this contain legal escalation threats?"),
}

evaluation = router.evaluate(
    state="Customer: Cancel my order #1082 and refund now, or I sue!",
    schema=schema,
)
print("Intent:", evaluation.results["intent"].selected)
print("Legal Threat:", evaluation.results["legal_threat"].value)

---

🔒 Cryptographic Checkpoint Provenance

{
  "checkpoint_file": "s1_decision_weights.pt",
  "checkpoint_sha256": "458c3b73157e05a99ae53b0a2e1eb090b3d434bc5facfb482b1fd884968f177d",
  "base_model": "Qwen/Qwen2.5-1.5B-Instruct",
  "conformal_q_hat": 0.95361328125,
  "nominal_coverage": 0.95,
  "calibration_samples": 10075,
  "test_samples": 8657,
  "license": "Apache-2.0"
}

📜 Citation

If you use Millennium-Jev in your research or production systems, please cite:

@software{millennium_jev_2026,
  author = {Ziqin Sun},
  title = {Millennium-Jev: Provably Safe Non-Autoregressive System 1 Decision Model},
  year = {2026},
  url = {https://github.com/sunziqin/millennium-jev}
}