ConceptNet Intent Classifier

Fine-tuned distilbert-base-multilingual-cased on the ConceptNet 4-layer enterprise voice intent taxonomy.

Accuracy

Standard test set: 100% (epochs 4 and 5)
Adversarial holdout: 99.3% (independently verified — Hugging Face community)
Fast-path classifier: 83% · <5ms latency
Dataset: 757 examples across 9 languages

Independent Evaluation

Independently tested by the Hugging Face community (john6666):

Confirmed 99.315% on reconstructed public test split
Grouped lexical-family holdout: 99.78%
Conclusion: "The obvious train/test leakage explanation did not survive that check"
All 4 identified improvements implemented within 24 hours

Cascade Performance

Threshold	Fast coverage	Fast accuracy	Final accuracy
0.50	69.2%	95.0%	95.9%
0.55	60.3%	98.9%	99.3%
0.65	43.2%	100%	100%

Layer Precedence

Mixed semantics: L4 > L3 > L2 > L1

The 4 Layers

L1 Basic — "Do X" — immediate execution
L2 Context-Aware — "Do X when Y" — conditional
L3 Predictive — "Do X before/ahead of/prior to Y" — proactive
L4 Autonomous — "Do X always" — persistent agent

Languages

English · French · Spanish · German · Italian · Portuguese · Chinese · Arabic · Russian

Usage

from transformers import pipeline
classifier = pipeline("text-classification", model="conceptnetUk/intent-classifier")
classifier("Send the report when the contract is signed")

conceptnetUk
/

intent-classifier