skychorus-sentiment-v1.0

SkyChorus Two-Stage Sentiment Model

This repository contains the frozen two-stage model used in the SkyChorus dissertation project.

The model is based on cardiffnlp/twitter-roberta-base-sentiment-latest and contains two independently trained classifiers:

• relevance_model/: classifies posts as relevant or irrelevant.
• sentiment_model/: classifies relevant posts as positive, neutral, or negative.

The models were trained on 1,000 manually annotated English Bluesky posts from five brands: Apple, Coca-Cola, Netflix, Nike, and Tesla.

Directory Structure

skychorus-sentiment-v1.0/
|-- relevance_model/
|-- sentiment_model/
`-- deployment_manifest.json
## Local Use

Download the model repository and configure SkyChorus with:

```text
SENTIMENT_MODEL_BACKEND=two_stage
TWO_STAGE_MODEL_DIR=models/skychorus-sentiment-v1.0

The TWO_STAGE_MODEL_DIR directory must directly contain the relevance_model/ and sentiment_model/ subdirectories.

Limitations

The model was trained and evaluated on a limited dataset of 1,000 posts from five brands. Predictions for other brands, time periods, and linguistic patterns should be interpreted with caution. The model is intended for exploratory research rather than representative market measurement.

Base Model

This model was fine-tuned from cardiffnlp/twitter-roberta-base-sentiment-latest.