Sentiment accuracy
—Waiting for resultsLSTM for the Win · Model results
Product review classification
A reproducible view of sentiment and product-topic classification on the newest controlled synthetic incoming batch. The dashboard reads the immutable run artifact generated by the model workflow.
RunLoading latest run…
Model executedLoading…
ScopeSynthetic · Incoming
Status Loading live data
Resolving the newest versioned workflow output from GitHub… View run files
Executive summary
Performance at a glance
Accuracy = correct predictions ÷ incoming reviews · 95% Wilson intervals shown below
Topic accuracy
—Waiting for resultsBoth labels correct
—Waiting for resultsIncoming reviews
—Unique incoming IDsBenchmark interpretation
Loading the latest model comparison…
The interpretation is calculated from the canonical workflow run.
Model diagnostics
Where the models perform
Accuracy
Model comparison
Recall by expected label
Sentiment class performance
Error pattern
Sentiment confusion matrix
Coverage
Topic volume and accuracy
Prediction behavior
Class balance and confidence
Sentiment counts
Expected vs. predicted
ExpectedPredicted
Average confidence
Certainty vs. accuracy
Confidence is the model probability assigned to its selected class. It is not the same as accuracy.
Review explorer
Inspect individual predictions
Loading reviews…
| ID | Review | Sentiment | Topic | Result |
|---|---|---|---|---|
| Loading model results… | ||||
Page 1
How to read this dashboard
Data provenance and limitations
Data. The primary cards and explorer display the newest synthetic incoming batch. Runs retain an immutable synthetic benchmark for longitudinal comparison. New runs also include a separately sourced real-world sentiment benchmark.
Versioning. latest.json resolves the newest retained run. Each run exposes one immutable run.json; paper analyses and any later CSV or Parquet views are derived directly from that same run artifact.
Continual learning. A controlled fraction of incoming reviews marked goldtest is promoted after evaluation. The synthetic benchmark and external benchmark never enter training.
Scope. Training, incoming evaluation and the longitudinal benchmark remain synthetic. Independent external validation uses the CC BY 4.0 UCI Sentiment Labelled Sentences Amazon subset and applies to sentiment only. The external source has binary positive/negative labels and no compatible product-topic labels, so it does not establish external topic generalization or full three-class sentiment coverage.