Scorers

All scorers implement cbtone.scorers.Scorer and return ToneScore (see Concepts). Build one with cbtone.scorers.get_scorer(name, **kw); heavy dependencies are imported lazily.

--scorer	Family	Needs	Languages	Notes
dict	Dictionary	core install	en	Phrase lexicon hawkdove_v1.csv shipped in the package, with a crude stemmer and negation handling. Written for cbtone following published construction principles; not a replication of any published list. Optional Loughran-McDonald features are downloaded at run time, never shipped.
fomc-roberta	Finance-tuned transformer	[transformers]; Hugging Face access to the gated model	en	gtfintechlab/FOMC-RoBERTa, revision pinned; CC-BY-NC-4.0, so non-commercial. Training data came from FOMC text, which matters for transfer to other banks. The paper uses a replication (fomc-roberta-repl) because access to the gated model was never granted: rebuild it with make fomc-roberta, or download it from thanatchon36/cbtone-fomc-roberta-repl (CC-BY-NC-4.0), then load it with --scorer ft --model-dir <dir>.
llm	Local LLM	a running Ollama and the model pulled	en, id, ms, th, zh	Default qwen3:4b, temperature 0, five prompt variants P1 to P5, optional --masked input and --evidence quotes. Calls are cached by content hash and resumable; the model tag and digest are recorded.
ft	Fine-tuned small model	[finetune] (training) or [transformers] (inference); --model-dir	the languages it was trained on	XLM-RoBERTa-base fine-tuned on LLM silver labels. It refuses training rows that come from non-commercial evaluation-only data.

The [api] extra is reserved for an optional hosted-API scorer; no such scorer is shipped yet, and none is ever required to reproduce results.

Contamination

A language model may have seen central-bank texts and the outcomes that followed. For that reason every LLM result is reported with:

• a masked variant (--masked): bank names, dates, numeric rate levels, currencies and country names replaced by placeholders;
• a post-cutoff split: documents dated after the model's documented training cutoff (or its public release date when none is documented).

A claim of predictive signal is made only where it survives both.

Choosing a scorer

There is no default winner: the pre-registered questions compare them. Read RESEARCH_QUESTIONS.md for what is tested, and DECISIONS.md for the model and prompt choices.

To write your own, see CONTRIBUTING.md.