banking77

Banking intent classification using an MLP baseline and parameter-efficient LoRA fine-tuning of RoBERTa.

Held-out intent benchmark on BANKING77 for few-shot intent classification, comparing TF-IDF, frozen MiniLM, and ProtoNet on unseen intents.

Benchmark and results for Closing the Compliance Gap: open-weights and proprietary LLMs for privacy-constrained contact center automation (under review, NeurIPS 2026 InfPriv workshop)

Fine-tuned DeBERTa-v3-base for 77-class BANKING77 intent classification with leakage-aware evaluation and statistical comparison against a TF-IDF baseline.

Fine-tunes Google FLAN-T5-base with LoRA (PEFT) on Banking77 to classify 77 banking intents. Trains only 1.77M params (0.71%) in ~2hrs on free Colab GPU. Served via FastAPI REST API with batch prediction, Docker support & evaluation metrics.

NLP lab comparing a Hybrid XGBoost against a fine-tuned DistilBERT for 77-way banking intent classification, evaluating F1, latency and structural trade-offs.

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.

Empirical study of candidate-order instability in non-autoregressive multi-candidate transformers, with cardinality scaling, permutation marginalization, cyclic shifts, calibration, and reproducible statistical evaluation.

A framework for benchmarking model reliability beyond clean accuracy using stability, robustness, severity-response, and paired evaluation.

Multi-agent customer message router with grounded RAG and deliberate refusal. Built on Banking77, evaluated with bootstrap confidence intervals.

BANKING77 intent classification: LoRA vs TF-IDF, held-out evaluation, and a guarded local API

Evidence-first BANKING77 intent routing with TF-IDF, RoBERTa, LoRA, calibration, uncertainty evaluation, PII redaction and governed FastAPI deployment.

Reproducible hill climbing for language-model classification

Risk-aware banking support router for 77 intents using calibrated classification, selective prediction, human escalation policies, and FastAPI.

A reproducible Apple Silicon learning lab: Qwen LoRA, BERT, DeBERTa, BANKING77, evaluation, failures, and a 94.12% classifier.

Verifiable handoffs between AI systems: Waggle packages decisions and evidence; Kea independently verifies, replays, or refuses them.