SignalLab.
A review-analysis workspace for understanding where sentiment models succeed, where they fail, and what synthetic augmentation really changes.
A historical product capture. Screens and functionality continue to evolve.
Understand the data before the model.
SignalLab is a local learning application for exploring customer reviews, training sentiment models and comparing their errors. The interface separates exploration, experiments, analysis, saved runs and prediction. Its small authored dataset demonstrates the workflow; it is not a benchmark of real-world model quality.
Make comparisons traceable.
Baseline experiments use TF-IDF and logistic regression. A separate PyTorch training workflow supports DistilBERT through Hugging Face. Analysis aligns predictions to exact review IDs, shows model disagreements, and reports performance by sentiment and review length with sample counts.
Keep the evidence, not just the score.
Completed baseline comparisons can be saved and reopened without retraining. Saved records include settings, provenance and prediction evidence. Detailed error analysis requires reattaching the matching dataset. Transformer experiment history remains a follow-up.
A question, not a claim.
The current augmentation baseline uses a deliberately weak prefix template. It does not demonstrate semantic synthetic-data generation or prove that augmentation improves models. Better controlled experiments, reviewed paraphrases and multiple seeds are learning goals.