SerbFin-Sentiment

First publicly available Serbian-language financial sentiment dataset. Short financial news headlines / lead sentences from Serbian business news sites, labeled for 3-class sentiment from an investor / market perspective (not emotional sentiment).

"Company cuts 5,000 jobs to boost margins" is positive from a market viewpoint even though it sounds bad. See the annotation guidelines below.

Sources & collection method

Collected from Serbian financial news sites: B92 Biz, BGDX Izdavaoci, Biznis.rs, Nova Ekonomija, eKapija Energija, eKapija Finansije, eKapija Industrija. Date range: 2016-04-01 to 2026-06-02.

Only short headline/lead text, source name, source URL, and date are stored -- no full article text is redistributed, per each site's terms of use.

Labeling

Three-stage weak-supervision -> human-gold pipeline (SPEC.md §4.4):

1. Silver labels -- Claude Haiku pre-labels the train pool (label_source: "llm-silver").
2. Gold test set -- 240 headlines, stratified sample, hand-labeled from scratch by a human annotator (label_source: "human"), never auto-labeled. The silver hint was hidden from the annotator until after each label was chosen (see data/ANNOTATION.md in the GitHub repo for full guidelines and the labeling tool used).

Test labels are 100% human -- this is what makes the "beats zero-shot LLM" comparison in the model card valid rather than circular (evaluating an LLM against its own silver labels would just measure self-agreement, not accuracy).

Label schema

label	meaning (investor / market perspective)
positive	Bullish -- good for company/sector/economy: profit up, beat expectations, new deal, growth, upgrade
negative	Bearish -- bad for company/sector/economy: loss, miss, layoffs, downgrade, default, legal penalty
neutral	No clear directional signal: appointment, schedule, mixed result, purely descriptive

Splits

split	n	source
train	351	llm-silver
validation	150	llm-silver
test	240	human (gold)

Class balance

train

label	count	%
positive	97	27.6%
negative	67	19.1%
neutral	187	53.3%

validation

label	count	%
positive	41	27.3%
negative	29	19.3%
neutral	80	53.3%

test (gold)

label	count	%
positive	80	33.3%
negative	81	33.8%
neutral	79	32.9%

Financial news skews neutral; no oversampling was applied to val/test (only ever appropriate for train, and not used here).

Inter-annotator agreement

Not computed for this release -- the gold test set was single-annotated. A second annotation pass (Cohen's kappa >= 0.6 target per SPEC.md §4.3) is left as future work.

License

CC BY 4.0. Headlines only, with source URLs included; no full articles are redistributed.

Intended use & limitations

Intended for training/evaluating Serbian financial sentiment models. Domain is short financial news headlines, not social media or emotional text -- performance on out-of-domain text (tweets, long-form articles, non-financial news) is untested. Class balance skews neutral; the positive/neutral boundary is the most common source of model confusion (see the companion model's confusion matrix at filipovicfilip222/serbfin-sentiment-llama-3.2-3b).

Companion model

filipovicfilip222/serbfin-sentiment-llama-3.2-3b -- QLoRA fine-tuned Llama 3.2 3B Instruct adapter trained on this dataset's train/validation splits and evaluated on test.