Quality Inspector
Automated AI evaluation, hallucination resistance scoring, semantic diversity, and benchmark decontamination
Quality Metrics Overview
Metrics & Grounding Controls
- Duplicate Detection4 levels (exact, fuzzy, embedding, MinHash LSH)
- Hallucination DetectionSource grounding + context citation matching
- Diversity Analysis5 dimensions (topic, source, instruction, response, domain)
- Semantic ScoringLLM Judge + heuristic fallback (perplexity & length)
- Human Review GateStratified sampling, approve, reject, edit in HITL queue
Dataset 5D Jury Certification & Fidelity Audit
Per-dataset evaluation across topic adherence, grounding, and safety dimensions
Adherence to user prompt & domain bounds
Factually grounded in domain sources
Syntactic structure & inquiry precision
Comprehensive statutory/technical depth
Jurisdiction & safety terminology adherence
Automated LLM-as-a-Judge Multi-Dimensional Rubric
Sample Quality Jury
Semantic Density & 2D Topic Coverage Visualizer
Topological Clusters (4)
GST & Tax Calculations
gst, calculate, tax
Statutory Deductions & Slabs
80c, deduction, epf
Compliance & Dispute Notices
notice, gstr-3b, penalty
Enterprise & MSME Registry
udyam, msme, aadhaar
Click any point on the canvas to inspect its instruction snippet and latent spatial coordinates.
Curriculum Complexity & Cognitive Tier Distribution
5-Tier Stratified BalancingFoundational Recall
Structured Logic
Multi-Step CoT
Tool Calling & Schema
Frontier Deduction
Evaluation Benchmark Decontamination Purity
100% Zero Test Leakage Certified (Audit Date: 26 Oct 2026)Overlap: 0.00%
Overlap: 0.00%
Overlap: 0.00%
Overlap: 0.00%
Overlap: 0.00%
1. 8-Gram MinHash LSH Index
All generated instructions and responses are broken into 8-token shingle n-grams and hashed across 128 permutation buckets to detect approximate overlaps.
2. Jaccard > 0.85 Hard Threshold
Any candidate sample exhibiting Jaccard coefficient exceeding 0.85 with benchmark test splits is automatically quarantined and purged from downstream training sets.
3. Zero Test Leakage Guarantee
Guarantees evaluated model checkpoints never memorize ground-truth benchmark prompts or test solutions, ensuring genuine zero-shot generalization.
Verified against evaluation splits across 5 frontier benchmarks. Zero ground truth contamination.