Executive Summary
ProcessVenue developed a high-consistency Hindi intent classification dataset to enable robust AI-assistant prompt routing. The project labelled 5,000 Hindi user prompts across six core assistant intents. Through dual independent annotation and structured verification, the dataset achieved 77.22% IAA, demonstrating dependable supervision quality and stable intent boundaries suitable for training production-grade intent routers.

Objective
Create a low-noise Hindi intent dataset aligned with real assistant traffic to improve routing reliability and supervised model learning for multi-intent and constraint-rich user prompts.

Key Challenges
Hindi prompts in assistant environments frequently contain overlapping or mixed intents, making classification non-trivial. Major complexity drivers included:
Multi-clause instructions combining tasks in one prompt
Constraint-heavy language (tone, format, tool usage, style requirements)
In-prompt domain switching (e.g., education → finance)
Code-switching between Hindi and English technical terms
Intent boundary overlap, especially: 1. Rewrite vs. Content Creation, 2. Explanation vs. Creative Writing

Maintaining labeling stability required strict guideline control and disciplined drift prevention.


Dataset Scope
The dataset reflects natural Hindi assistant usage patterns:
Conversational, instruction-style prompts
Broad domain distribution (education, healthcare, economics, environment, etc.)
High density of real routing constraints
Balanced mix of short commands and long multi-clause prompts
These attributes support training routers that generalize to real production behaviour.


Methodology: Double Blind
ProcessVenue employed a three-layer annotation workflow to reduce noise and ensure consistent intent boundaries.

Annotator1: primary label assignment
Annotator2: independent secondary label
Verifier: resolves conflicts, finalizes labels, logs edge cases
IAA computed post-reconciliation to confirm boundary stability