daudaudinang/edgebench-bilingual-jev

EdgeBench-JEV: Decision & Bilingual Reasoning Benchmark

EdgeBench-JEV is a comprehensive evaluation benchmark designed to measure the real-world operational decision-making, latency scaling, and calibration of LLMs and Decision Models (such as Quyet-1.0, Clef-27B, H2O-Danube).

It focuses on authentic enterprise scenarios across Southeast Asia (banking, e-commerce customer support, port logistics, supply chain, insurance, and healthcare).

---

🗂️ Dataset Overview

File	Size	Records	Description
bilingual_bench.jsonl	7.7 MB	1,000 cases	Enterprise operational reasoning across 5 domains, bilingual Vietnamese-English with Chinese trade logistics.
bilingual_bench.rationale.jsonl	192 KB	1,000 cases	Ground-truth step-by-step reasoning rationales for every case.
context_scaling_100.jsonl	641 KB	100 cases	Latency and throughput scaling suite across 6 context tiers (128t, 256t, 512t, 1024t, 2048t, 4096t).
context_scaling_300.jsonl	1.9 MB	300 cases	Concurrency scaling benchmark under variable context lengths.
jev_standard_231.jsonl	587 KB	231 cases	Standard JevBench decision benchmark suite for strict schema compliance and parity testing.

---

🏢 Bilingual Benchmark Breakdown (1,000 Cases)

1. Categories (200 cases each)

• TIME: Timestamp extraction, international flight timezone offset calculations, estimated arrival time (ETA), maritime shipping cutoff times, and SLA breach deadlines.
• CALC: Multi-step arithmetic reasoning, tiered discounts, logistics shipping coupons, cashback calculations, and VAT/surcharge calculations.
• LABEL: Multi-dimensional sentiment analysis, ticket priority classification (P0 Life-safety, P1 Urgent, P2 Normal), and customer churn prediction.
• NOISY: Ultra-long input contexts (10,000–15,800 characters) containing conflicting policy amendments, superseded terms, system logs, and spam email noise (Needle in a Haystack).
• MIXED: Natural code-switching conversations (Vietnamese-English, Vietnamese-Chinese) and specialized international trade terminology (Bill of Lading, demurrage, COD).

2. Languages

• Vietnamese (vi): 500 cases
• English (en): 500 cases
• Chinese elements (has_chinese): 206 cases (shipping manifests, foreign trade names, bilingual invoices)

3. JEV Decision Primitives

• choice (340 cases): Select exact categorical label from a finite criteria set.
• noul (330 cases): Binary decision (True / False / Refusal).
• score (330 cases): Ordered rating scale (1–5 Likert or risk tiers).

---

📋 Data Schema

Each sample in bilingual_bench.jsonl follows this schema:

{
  "id": "case-time-001-vi",
  "category": "time",
  "language": "vi",
  "state": {
    "now": "2026-10-15 08:30 (Thứ Năm), Asia/Ho_Chi_Minh",
    "channel": "Zalo CSKH GiaoHangHanh",
    "participants": [{"id": "u1", "role": "customer", "name": "Anh Tuấn"}],
    "order": {"id": "GHN-881920", "created_at": "2026-10-14 14:00", "sla_hours": 24}
  },
  "questions": {
    "sla_breach": {
      "type": "noul",
      "instructions": "Đơn hàng GHN-881920 có bị vi phạm cam kết SLA 24h nếu giao lúc 15:30 chiều nay không?"
    }
  },
  "expected": {
    "sla_breach": true
  }
}

---

📊 Benchmark Results on Tested Models

Model	JevBench (231) Accuracy	Bilingual (1,000) Accuracy	Throughput (req/s)	VRAM Footprint
Quyet-1.0-Large-Mixed-NVFP4	89.61%	86.70%	0.68	20.5 GB
Quyet-1.0-Large (BF16)	90.91%	86.20%	0.21	71.8 GB
Clef-27B (BF16)	88.31%	85.10%	0.24	58.2 GB
H2O-Danube-4B (BF16)	71.43%	68.40%	1.15	8.9 GB

Evaluated on NVIDIA GB10 (128GB Unified Memory, Blackwell sm_121 architecture).

---

Citation & License

Apache-2.0 License.

@dataset{edgebench_jev_2026,
  title  = {EdgeBench-JEV: Enterprise Operational Decision and Bilingual Reasoning Benchmark},
  author = {VCCorp AI Team},
  year   = {2026},
  url    = {https://huggingface.co/datasets/daudaudinang/edgebench-bilingual-jev}
}