Rebenchmarking Financial PhraseBank: Fixed Splits, Decontaminated Baselines, and Augmentation That Does Not Replicate
Ankit Aglawe · In preparation. Experiments complete; writing in progress.
What this paper is
Financial PhraseBank is the de facto benchmark for financial sentiment, yet published numbers on it are largely incomparable: no official split exists, the most-used baseline was evaluated on its own training data, and augmentation gains reported on private splits do not survive a controlled one. This paper releases a fixed stratified split with evaluation scripts, measures the contamination inflation of the standard baseline under an identical harness, and reports a six-arm test in which both LLM-synthetic and volume augmentation lose to training on the clean data alone.
Status
In preparation. Experiments complete; writing in progress. No DOI exists yet; this page is the paper's permanent home and will carry the publication record when it ships. The experiments and artifacts behind it are public as they land.