LaelaZ/distilbert-emotion

Text Classification • 67M • Updated 4 days ago • 71

Note Flagship — DistilBERT fine-tuned on dair-ai/emotion. Honest held-out eval: accuracy 0.920, macro-F1 0.874, with confusion matrix + error report.

Note Interactive demo of the fine-tuned 6-emotion classifier.

Note Production layer: typed API, dynamic batching, Prometheus + Grafana, load tests — serves the real model.

Note 2,829 synthetic instruction prompts across 8 domains — generated AND quality-graded by SynthKit (grade B, validity 100).

Note SynthKit demo set with a built-in train/eval contamination example (5 deliberate overlaps).

Note 5-table synthetic e-commerce warehouse (~104k rows) for analytics & tabular ML.

Note SynthKit — generate synthetic data AND grade it (validity / uniqueness / diversity / contamination).

Note Grounded legal-info Q&A (IL divorce): retrieval + scope gate + citations, with an anti-hallucination eval harness.

Note Model/data evaluation toolkit.

Note LLM red-teaming: prompt-injection / jailbreak / leakage probes with severity + compliance mapping.

Note SWC-mapped Solidity static analyzer with machine-readable reports for CI gating.

Note Pairwise RLHF preference-rating tool.

Note QC/validation harness for agentic-workflow scenarios.

Note SCORM e-learning package QA validator.

Note SupportCopilot — RAG e-commerce support agent; deflection rate measured and translated to agent-hours saved.

Note InterviewCoach — JD-aware mock interviews with rubric-scored feedback.

Note CommercePipeline — DuckDB warehouse, staged SQL (staging→marts), data-quality gates, Streamlit dashboard.

Note CNN image classifier with Grad-CAM explainability.

Note LSTM time-series forecasting demo.

Note Neural network implemented from scratch (educational).

Note Citation-grounded Illinois divorce corpus, 33 passages covering the full process (grounds, property, support, children, forms, service, remote hearings) for RAG and legal-NLP.