Wadden/sentiment-versioning-demo

Versioned Sentiment Classifier

Three traceable DistilBERT variants from the model-versioning lab. Each variant is a Hub commit and tag; the best one is exposed through the production branch for independent deployment and rollback.

Versions

Tag	Training data	Max length
v1	2,000 Rotten Tomatoes reviews	128
v2	2,000 Rotten Tomatoes reviews	256
v3	2,000 IMDb reviews	256

Common-test comparison

All variants were evaluated on the same Rotten Tomatoes test split.

Version	Accuracy
v1	0.7955
v2	0.8011
v3	0.7955

Promoted version: v2 (branch production).

Usage

from transformers import pipeline
classifier = pipeline("sentiment-analysis", model="Wadden/sentiment-versioning-demo",
                      revision="production")

Interactive production demo

The quantized ONNX export of the promoted v2 revision is executed directly in the browser:

Open the production Space