thanatchon36/cbtone-silver-sentences

cbtone silver sentences

The 3,755 central-bank sentences used to train thanatchon36/cbtone-ft-xlmr, each with a hawkish / dovish / neutral label assigned by a local LLM (qwen3:4b, registered prompt P1@8bfb4852, temperature 0). Companion to the paper Reading the Room: Do Local-LLM Tone Scores of Central-Bank Communication Carry Policy Signal? (Chongmankhong, 2026).

📄 Paper (web) · 💻 Code: cbtone · 🔖 DOI 10.5281/zenodo.23371641

Silver labels, not gold. Every label comes from an LLM, not from a human annotator. On the human-labelled FOMC test split the same LLM and prompt reach a three-class macro-F1 of 0.48. Treat the labels as a noisy reading, suitable for distillation and for studying LLM behaviour, not as ground truth.

Two configurations, because not every source may be redistributed

Config	Rows	Text included	Sources
public (default)	1,187	✅	Federal Reserve (public domain), Reserve Bank of Australia (CC BY 4.0), ECB monetary-policy press releases (reproduction permitted with acknowledgement)
ids_only	2,568	❌	Bank of England, Monetary Authority of Singapore, Bank Indonesia, Bank Negara Malaysia, ECB press conferences

The second group's terms do not permit redistributing their text, so ids_only gives the document URL, the document's SHA-256, the sentence position and a SHA-256 of the sentence text, plus the label. With the cbtone library (cbtone fetch then cbtone snapshot) the text is rebuilt locally from the banks' own websites and can be checked against sentence_sha256.

33 further training sentences could not be traced back to a single source document and are omitted, so the dataset has 3,755 of the model's 3,788 training sentences.

from datasets import load_dataset

public = load_dataset("thanatchon36/cbtone-silver-sentences", "public", split="train")
ids = load_dataset("thanatchon36/cbtone-silver-sentences", "ids_only", split="train")

Fields

Field	Description
bank	fed, ecb, rba (public); boe, mas, bi, bnm, ecb (ids_only)
doc_id	bank:date:doc_type:lang, the cbtone document identifier
date	Release date of the source document
doc_type	statement, minutes, press_conference or mps (MAS Monetary Policy Statement)
para_idx, sent_idx	Position of the sentence in the document's tagged policy text
lang	en, id or ms
label	hawkish, dovish or neutral (LLM silver label)
text	Sentence text (public only)
url	Source document URL
doc_sha256	SHA-256 of the source document as fetched
source_licence	Terms of the source (public only)
sentence_sha256	SHA-256 of the UTF-8 sentence text (ids_only only), to verify a local rebuild

Composition

Bank	public	ids_only
Federal Reserve	374	—
Reserve Bank of Australia	681	—
European Central Bank	132	67
Bank of England	—	741
Monetary Authority of Singapore	—	666
Bank Indonesia	—	671
Bank Negara Malaysia	—	423

Labels over all 3,788 training sentences: neutral 2,890 · dovish 546 · hawkish 352. Languages: English 3,658 · Indonesian 66 · Malay 64.

How it was built

Sentences come from the policy paragraphs (decision, forward guidance, outlook, risks) of each bank's primary decision documents, fetched politely (robots.txt, ≥ 2 s per request, cached with SHA-256) by cbtone. The FOMC hawkish/dovish dataset (CC BY-NC) is not included and was not used for training; the 200 Southeast Asian sentences reserved for evaluation are excluded. The export script is experiments/export_hf.py.

Licence

Per source, as listed above and in docs/sources/LICENSES.md. The labels and the metadata are released under CC BY 4.0. Cite the source institution when reusing its text.

Citation

Please cite the paper; if you use the cbtone library, also cite the software.

@misc{chongmankhong2026reading,
  author = {Chongmankhong, Thanatchon},
  title  = {Reading the Room: Do Local-{LLM} Tone Scores of Central-Bank Communication Carry Policy Signal? Pre-registered Evidence from Seven Central Banks in Three Languages},
  year   = {2026},
  note   = {Preprint},
  url    = {https://thanatchon36.github.io/cbtone/paper/}
}

@software{chongmankhong_cbtone,
  author  = {Chongmankhong, Thanatchon},
  title   = {cbtone},
  year    = {2026},
  version = {0.1.3},
  url     = {https://github.com/thanatchon36/cbtone},
  doi     = {10.5281/zenodo.23371641}
}

Disclaimer. Independent open-source project. Public documents only; represents the views of no institution; not investment advice.