DeBERTa-v3 multi-task financial sentiment: beats a public baseline on 2/3 benchmarks, honest about where it doesn't
Fine-tuned microsoft/deberta-v3-base jointly on three financial sentiment datasets (Financial PhraseBank, Twitter Financial News, FiQA) -- one shared backbone, two heads (3-class classification + continuous regression).
Beats the current best public multi-dataset financial sentiment baseline (pmatorras/financial-sentiment-analysis) on Twitter (+6.4 points accuracy) and FiQA (+4.8 points accuracy). Falls short on Financial PhraseBank (87.1% vs 95.9%) -- the model card has a confusion matrix showing the errors are almost all at the neutral/positive boundary, not random noise, plus the working theory for why (FinBERT's domain-specific pretraining likely helps on PhraseBank's formal, 2014-era news style specifically).
Also reports bootstrap confidence intervals and a continuous MAE/MSE baseline for FiQA that I couldn't find in a comparable public model.
Full training and evaluation notebook (real outputs baked in, run end-to-end on a single T4) is on GitHub: https://github.com/ProPrak01/deberta-v3-financial-sentiment
Feedback and questions welcome, especially on the PhraseBank gap -- happy to hear if anyone's tried a similar setup and seen something different.