Roadmap and known gaps¶

An honest list of what is missing. Items are not promises or dates.

Known gaps in the evidence¶

One evaluation set (884 test sentences, one source, one annotator). Inter-annotator agreement is unmeasured, and the neutral class has only a working definition (normal, everyday conversation), not a written guideline.
No out-of-domain tests: social media, long text, code-mixed text, other scripts, dialects, sarcasm.
No significance shown for XLM-R over the baseline (p = 0.116).
Synthetic data's value is unproven: a 0.02 validation gain without a significance test.
Overconfident probabilities with no calibration layer.

The Docker files, and the API tests that need FastAPI and the trainer tests that need PyTorch, run in CI but were not run while writing this documentation.
No authentication or rate limiting in the API.
No streaming or async batching; one request is processed at a time.
No model quantization or ONNX export for faster CPU use.
No calibration tool (temperature scaling).
No published baseline model on the Hugging Face Hub.
No Hugging Face pipeline integration for the cleaned text.
The documentation theme, Material for MkDocs, announces its end of life on 5 May 2027 (security fixes continue until then). Its maintained successor is Zensical; the site builds fine today, and a move can wait.

A second annotator and an agreement study (Cohen's kappa).
A larger and more varied test set.
Explanations for individual predictions (for the baseline, the strongest words).
A distilled or quantized model for CPUs.
Sentence splitting for longer documents.
More classes or aspect-level sentiment.

Corrections, new labeled data with a clear license, and evaluation on other domains are especially valuable. See Contributing.