clinc150

Code Implementation of TDS Article "Semi-supervised Intent Classification with GAN-BERT"

Student project regarding out-of-domain text classification methods comparison on CLINC150 dataset.

Cost-aware hybrid router for multi-agent LLM systems: keyword → embedding → LLM cascade. Matches LLM-only accuracy (82.6% vs 82.9%, McNemar p>0.3) with 74% fewer LLM calls on CLINC150.

Finite-sample guarantees for Jev (TypeSafe's System One). Conformal risk control turns calibrated probabilities into certified routing thresholds; prediction-powered inference audits them. 2,412 decisions on CLINC150 for $0.23 — including the shift and prevalence cases where the guarantee breaks.

Reproducible Laya-CoreML vs Jev benchmark for zero-shot intent classification on Banking77, ArBanking77, and CLINC150. Includes paired accuracy and per-dataset results.

Frozen encoder + Mahalanobis prototype for class-incremental intent classification. 50+ experiments across BANKING77, CLINC150, HWU64, AG News. Matches fine-tuned baselines at 5MB state with zero forgetting, order-invariance, 455 QPS.

Voice-enabled chatbot that accepts spoken input: Whisper transcribes the speech, a fine-tuned DistilBERT model classifies it into one of 41 CLINC150 intents, and a response is returned. 0.9473 test accuracy, 0.039 word error rate. FastAPI backend on Railway, React front end on Vercel.