remehostingservices/finance-news-sentiment-35k
Finance News Sentiment 40k 39,965 English financial news headlines, collected from public Telegram finance news-wire channels, labeled for 3-class sentiment (positive / negative / neutral) and a secondary topic label, by two independent LLM judges from different model families with an arbiter settling disputes. A FinBERT model fine-tuned on this data reaches test accuracy 0.847 / macro F1 0.810: remehostingservices/finbert-finance-news-sentiment. Code, training scripts and the… See the full description on the dataset page: https://huggingface.co/datasets/remehostingservices/finance-news-sentiment-35k.
Dataset card: add label reliability audit section (link to nlp-judge-audit)
v2: 39,965 rows (+4,997 active-learning rows, same rulebook and judges); license clarified (annotations CC BY-NC 4.0, headline texts remain their sources')
Rename to Finance News Sentiment 35k; add GitHub link
Rename to Finance News Sentiment 35k; add GitHub link
Card: list source channels with row counts and date ranges
Card: judge Claude Opus 5, arbiter Claude Fable 5
Card: name the arbiter model, drop personal name from citation
Add dataset splits and card
initial commit