IndicJevBench

The first open benchmark for structured decision AI in Indian languages.

IndicJevBench tests whether a model can answer typed questions about a customer message and return calibrated probabilities. Questions cover intent classification, urgency scoring, escalation detection, and agent routing.

Languages covered: Hindi (Devanagari), Hindi (Romanized / Hinglish), Bengali, Tamil, Telugu, Kannada, Malayalam.

---

Tasks

Dataset	Type	Languages	Items	License
intent_massive	Intent classification	hi, bn, ta, te, kn, ml	57,922	CC BY 4.0
fintech_banking77	Banking intent	hi, bn, ta, te, kn, ml	7,816	CC BY 4.0
hinglish_lid	Script detection (Devanagari / Latin / both)	hi-Latn (Hinglish)	3,220	CC BY 4.0
synthetic_enterprise	Urgency, escalation, routing	hi, hi-Latn	444	CC BY 4.0

Total: 69,402 items.

---

Format

Each item looks like this:

{
  "id": "massive-hi-Deva-0001-q0",
  "family": "intent",
  "lang": "hi-Deva",
  "state": "मुझे अपनी फ्लाइट बुक करनी है",
  "question": {
    "type": "choice",
    "instructions": "What is the customer's intent?",
    "options": ["book_flight", "cancel_flight", "check_status", "other"]
  },
  "expected": 0,
  "license": "CC BY 4.0"
}

Question types:

• choice: pick one option from a list
• score: rate on an ordered scale (urgency 1 to 5)
• noul: yes or no decision (should this be escalated?)

---

Scoring

Models are scored on the IndicJevScore, a single number from 0 to 100:

Axis	Weight
Accuracy	35%
Calibration (ECE + Brier)	25%
Speed	20%
Cost per 1k decisions	20%

The score is the geometric mean of all four axes.

---

Sources

• MASSIVE (Amazon, CC BY 4.0) - intent classification across Indic languages
• Banking77 (PolyAI, CC BY 4.0) - banking intent, translated to Indic via NLLB-200
• COMI-LINGUA (IIT Gandhinagar, CC BY 4.0) - Hinglish source text; reshaped in v1.1 to script-detection MCQ derived from Unicode script presence in the state text
• Synthetic (cmul8, CC BY 4.0) - enterprise support data generated for UPI payments, e-commerce, and agent routing

---

Leaderboard

Measured by cmul8 on 2026-09-22 using 1000 items per dataset (where available). hinglish_lid re-measured against v1.1.0 reshape on 2026-09-24.

Model	Type	fintech_banking77	synthetic_enterprise	intent_massive	hinglish_lid (v1.1)
OpenJev (Qwen3.5-4B)	local	acc=0.620 / IJScore=78.3	acc=0.606 / IJScore=79.6	N/A (>16-option limit)	acc=0.555 / IJScore=73.0
Laya-multilingual	local	acc=0.455 / IJScore=69.3	acc=0.351 / IJScore=58.7	acc=0.350 / IJScore=61.7	acc=0.080 / IJScore=30.2
GPT-4o-mini (OpenRouter)	API	acc=0.775 / IJScore=56.1	acc=0.670 / IJScore=46.6	acc=0.580 / IJScore=51.3	(v1.0 — re-measuring)
Laya (base)	local	acc=0.256 / IJScore=56.2	acc=0.369 / IJScore=64.5	acc=0.199 / IJScore=50.2	acc=0.334 / IJScore=64.2
Qwen3-4B (zero-shot logprob)	local	acc=0.373 / IJScore=55.3	acc=0.525 / IJScore=68.5	acc=0.316 / IJScore=49.6	acc=0.598 / IJScore=72.9

Notes:

• OpenJev has a hard limit of 16 options; intent_massive (up to 60 options) cannot be evaluated with OpenJev.
• The primary comparison tasks are fintech_banking77, intent_massive, and synthetic_enterprise.
• hinglish_lid (v1.1): majority-class baseline is 0.635 (code-mixed). No baseline listed above clears it — best is Qwen3-4B at 0.598. This gap is the reason we augmented Nirṇaya's training set with 5k script-detection examples (see nirnaya v1.1 continuation training).
• GPT-4o-mini re-measurement on v1.1 hinglish_lid pending (v1.0 run had 46/200 JSON parse errors and was invalid).

---

Changelog

v1.1.0 (2026-09-23) — hinglish_lid reshape

hinglish_lid in v1.0 was packaged directly from COMI-LINGUA, whose options field contains 4–16 stringified per-word LID annotations (~1.3k chars each, up to ~15k chars per item). The natural-language instruction asked "which language/script is used?" but the options answered a different question ("which per-word tagging is correct?") and were not learnable in that shape. In v1.1.0 we keep the same 3,220 source states and instruction stem but replace the options with a three-way script-detection MCQ (Devanagari-only / Latin-only / both), with the gold label derived from Unicode script presence in the state text.

• Same file path: v1/hinglish_lid.jsonl
• Same item count: 3,220
• New gold distribution: 2,046 code-mixed / 899 Latin-only / 275 Devanagari-only (majority-class baseline ≈ 63.5%)
• All other splits unchanged.

---

Running the Benchmark

The benchmark harness and full code are coming soon.

---

License

This dataset is released under CC BY 4.0. Individual task licenses are listed in the table above.

Built by cmul8. Inspired by JevBench by Florian Standhartinger.