🐂 FinSense distilbert v2 — financial news sentiment, tiny and fast v2 — retrained on a cleaner recipe with a published split. Same 67M parameter speed your pipelines already rely on, more accurate than v1 on a properly held out benchmark. positive / neutral / negative for headlines, news wires, analyst sentences. Built on ModernBERT base — Flash Attention fast, 149M params, runs happily on CPU. Benchmarks Financial PhraseBank (the standard benchmark for this task), held out test set, identical harness for every row: Model Accuracy Macro F1 🐂 This model (v2) 0.8447 0.8316 v1 (previous weights) 0.8323 0.8064 FinBERT (reproducible benchmark¹) 0.8423 0.8439 +1.2 accuracy / +2.5 F1 over v1 — and it now edges past FinBERT's reproducible benchmark at a third of the size.¹ Want maximum accuracy? The ModernBERT flagship scores 0.8675. ¹ Independently replicated score of the public FinBERT checkpoint (Thomas, 2024). FinBERT scores higher (0.88) when evaluated on FPB samples overlapping its own training data; FinSense's test set is fully held out. Split script + raw eval outputs ship in this repo. Labels id label example 0 negative "Operating profit fell to EUR 35.4 mn from EUR 68.8 mn." 1…
We use cookies for essential functionality and analytics. You can accept or reject analytics cookies.Cookie policy