Words That Unite The World: A Unified Framework for Deciphering Central‑Bank Communications Globally

Authors

Role	Contributors
*Equal‑first ()**	Agam Shah, Siddhant Sukhani, Huzaifa Pardawala*
Core contributors (†)	Saketh Budideti†, Riya Bhadani†, Rudra Gopal†, Siddhartha Somani†, Michael Galarnyk†, Rutwik Routu†, Soungmin Lee†
Additional contributors	Akshar Ravichandran, Eric Kim, Pranav Aluru, Joshua Zhang, Sebastian Jaskowski, Veer Guda, Meghaj Tarte, Liqin Ye, Spencer Gosden, Rachel Yuh, Arnav Hiray, Sloka Chava, Sahasra Chava, Dylan Kelly, Aiden Chiang, Harsit Mittal, Sudheer Chava

🐝 Georgia Institute of Technology
Contact: {ashah482, ssukhani3, hpardawala3}@gatech.edu
Links:
- 🌐 WCB Website
- 🤗 Data & Models
- 💻 Code

Note – * denotes equal first authors, † denotes core contributors.

Paper & Website

🌐 Explore interactive visualisations on the WCB website

Dataset Overview

Dataset	Value
Central Banks	25
Years	1996 – 2024
Scraped Sentences	380,200
Annotated Sentences	25,000
Total Words	10,289,163
Corpus Size (tokens)	2,661,400
Sentences / Year*	13,110.34
Words / Sentence*	27.06

Model Information

Models	Value
Pre‑trained Language Models	7
Large Language Models	9
Best Stance Model*	`RoBERTa‑Large` (0.740)
Best Temporal Model*	`RoBERTa‑Base` (0.868)
Best Uncertainty Model*	`RoBERTa‑Large` (0.846)
Benchmarking Experiments	15,075
Few‑shot	✅
Few‑shot + Ann. Guide	✅

Annotation Details

Annotations	Value
Annotators	104
Annotation Guides	26
Annotation Steps	6
Tasks
Stance Detection	Hawkish, Dovish, Neutral, Irrelevant
Temporal Classification	(Not) Forward‑looking
Uncertainty Estimation	(Un)certain

Figure 1. A summary of the World Central Bank (WCB) dataset and experiments.
We systematically scrape, clean, and analyze 1996 – 2024 communications from 25 central banks at the sentence level, yielding 380,200 sentences (avg. 27.06 words/sentence). An annotated subset of 25,000 sentences spans three tasks (Stance Detection, Temporal Classification, and uncertainty Estimation) using comprehensive individual annotation guides and detailed instructions for annotation. We benchmark seven PLMs and eight LLMs on these tasks, under a bank-specific (1,000 bank-specific annotated sentences) and global setup (25,000 annotated sentences). The performance of the General (All-Banks) Setup model for each task is showcased in the figure.

In these tables, * represents that it is an average.

Abstract

Central banks around the world play a crucial role in maintaining economic stability. Deciphering policy implications in their communications is essential, especially as misinterpretations can disproportionately impact vulnerable populations. To address this, we introduce the World Central Banks (WCB) dataset, the most comprehensive monetary policy corpus to date, comprising over 380k sentences from 25 central banks across diverse geographic regions, spanning 28 years of historical data. After uniformly sampling 1k sentences per bank (25k total) across all available years, we annotate and review each sentence using dual annotators, disagreement resolutions, and secondary expert reviews. We define three tasks: Stance Detection, Temporal Classification, and Uncertainty Estimation, with each sentence annotated for all three. We benchmark seven Pretrained Language Models (PLMs) and nine Large Language Models (LLMs) (Zero-Shot, Few-Shot, and with annotation guide) over 15,075 experiments. We find that a model trained on aggregated data across banks significantly surpasses a model trained on an individual bank's data, confirming the principle "the whole is greater than the sum of its parts". Additionally, rigorous human evaluations, error analyses, and predictive tasks validate our framework's economic utility. Our artifacts are accessible through the WCB Homepage, HuggingFace, and GitHub under the CC-BY-NC-SA 4.0 license.

Repository Layout

Path	Description
`cleaned_data/`	Markdown and txt files with cleaned data
`configs/`	configs for the LLM experiments
`croissant_files/`	croissant files for the 25k annotated sentences and the full corpus (380k sentences)
`final_data/`	csv files with annotated sentences
`llm_inference_outputs/`	files containing outputs of the LLMs for different experiments
`llm_inference_logs/`	logs for the LLM experiments
`raw_data/`	Raw documents (PDF, txt, docx)
`resources/`	Figures & logos for the paper/README
`sanitized_data/`	txt files with only sentences
`src/llm_benchmarking/`	LLM experimentation pipeline
`src/plm_benchmarking/`	Pre‑trained encoder benchmarks
`src/additional_experiments/`	Scripts for generating synthetic meeting minutes, performing ablation studies, etc.
`utils/`	Utility scripts for logging and other helper functions.
`master_file_metadata.json`	Metadata file containing detailed information about the dataset (JSON).
`master_file_metadata.xlsx`	Metadata file containing detailed information about the dataset (Excel).

Format of Released Data

Released Splits

Per‑bank datasets: gtfintechlab/<bank_slug>, 3 seeds each.
Aggregated dataset: gtfintechlab/all_annotated_sentences_25000.

Metadata

The master_file_metadata.json in the root directory contains the metadata for the entire dataset. The entries are formatted as follows:

    { 
      "central_bank_name": { 
        "year": { 
          "document_key": {
            "release_date": "DD-MM-YYYY",
            "start_date": "DD-MM-YYYY",
            "end_date": "DD-MM-YYYY",
            "minutes_link": "URL to the source document",
            "cleaned_document_name": "Filename of cleaned document",
            "original_document_name": "Filename of original document",
            "sentences": [
              "First sentence from the document.",
              "Second sentence from the document.",
              "..."
            ]
          }
        }
      }
    }

Environment Setup

Tested on Python 3.9 – 3.11.
GPU support: CUDA 11.8 (optional).

Clone and create a virtual environment

git clone https://github.com/gtfintechlab/WorldCentralBanks.git
cd WorldCentralBanks
python -m venv .venv              # or: conda create -n wcb python=3.10
source .venv/bin/activate         # # conda activate wcb

Install Dependencies

pip install -r requirements.txt

Configure API Keys

Copy the sample file:

cp .env.example .env

Open .env and paste your credentials

HF_TOKEN= # read or read‑write scope

TOGETHERAI_API_KEY=

OPENAI_API_KEY=

GEMINI_API_KEY=

Code - getting started

Note: Ensure your environment is correctly set up before running the scripts.

Bank Dictionary: Mapping Dataset Keys to Official Central Bank Names

  {
    "bank_negara_malaysia": "Bank Negara Malaysia",
    "bank_of_canada": "Bank of Canada",
    "bank_of_the_republic_colombia": "Bank of the Republic (Colombia)",
    "bank_of_england": "Bank of England",
    "bank_of_israel": "Bank of Israel",
    "bank_of_japan": "Bank of Japan",
    "bank_of_korea": "Bank of Korea",
    "bank_of_mexico": "Bank of Mexico",
    "central_bank_of_the_philippines": "Central Bank of the Philippines",
    "bank_of_thailand": "Bank of Thailand",
    "central_bank_of_brazil": "Central Bank of Brazil",
    "central_bank_of_chile": "Central Bank of Chile",
    "central_bank_of_egypt": "Central Bank of Egypt",
    "central_bank_of_the_russian_federation": "Central Bank of the Russian Federation",
    "central_bank_republic_of_turkey": "Central Bank of Turkey",
    "central_bank_of_china_taiwan": "Central Bank of China (Taiwan)",
    "central_reserve_bank_of_peru": "Central Reserve Bank of Peru",
    "european_central_bank": "European Central Bank",
    "federal_reserve_system": "Federal Reserve System",
    "monetary_authority_of_singapore": "Monetary Authority of Singapore",
    "national_bank_of_poland": "National Bank of Poland",
    "peoples_bank_of_china": "People's Bank of China",
    "reserve_bank_of_india": "Reserve Bank of India",
    "reserve_bank_of_australia": "Reserve Bank of Australia",
    "swiss_national_bank": "Swiss National Bank"
}

Dataset Availability

Access the Dataset on Hugging Face

Access the 25k annotated sentences dataset on Hugging Face

The World Central Bank (WCB) dataset is available on Hugging Face. It includes:

Individual central banks' datasets
Aggregated dataset (25k annotated sentences, 1k per bank)
Full corpus (380,200 sentences without splits)

Loading the Full Corpus (380k Sentences)

from datasets import load_dataset

dataset = load_dataset("gtfintechlab/WCB_380k_sentences")

Loading the Aggregated Dataset (25k Annotated Sentences)

from datasets import load_dataset

dataset = load_dataset("gtfintechlab/all_annotated_sentences_25000", '{SEED}')

Loading a Specific Central Bank's Dataset

from datasets import load_dataset

dataset = load_dataset("gtfintechlab/{bank_name}", '{SEED}')

Available Seeds

5768
78516
944601

Croissant Files

Full corpus (380k sentences): root/croissant_files/croissant_WCB_380k_sentences.json
Annotated dataset (25k sentences): root/croissant_files/croissant_all_annotated_sentences_25000.json

Loading Models from Hugging Face

Using Pre-Trained Models for WCB Tasks

Use bank="WCB" for the best-performing general model, or any key from bank_map for a bank‑specific model.

Below are examples for all three tasks:

1. Stance Detection

Model Name Pattern: model_<bank>_stance_label
Labels: Hawkish, Dovish, Neutral, Irrelevant

Intended Use
This model is designed for researchers and practitioners working on subjective text classification, particularly within financial contexts. It assesses the Stance attribute, aiding in the analysis of subjective content in financial communications.

How to Use

from transformers import pipeline, AutoTokenizer, AutoModelForSequenceClassification, AutoConfig

# Load tokenizer, model, and configuration
tokenizer = AutoTokenizer.from_pretrained("gtfintechlab/model_{bank}_stance_label", do_lower_case=True, do_basic_tokenize=True)
model = AutoModelForSequenceClassification.from_pretrained("gtfintechlab/model_{bank}_stance_label", num_labels=4)
config = AutoConfig.from_pretrained("gtfintechlab/model_{bank}_stance_label")

# Initialize text classification pipeline
classifier = pipeline('text-classification', model=model, tokenizer=tokenizer, config=config, framework="pt")

# Classify Stance
sentences = [
  "[Sentence 1]",
  "[Sentence 2]"
]
results = classifier(sentences, batch_size=128, truncation="only_first")
print(results)

Label Interpretation

LABEL_0: Neutral; neither hawkish nor dovish, or both sentiments present.
LABEL_1: Hawkish; supports contractionary monetary policy.
LABEL_2: Dovish; supports expansionary monetary policy.
LABEL_3: Irrelevant; unrelated to monetary policy.

2. Temporal Classification

Model Name Pattern: model_<bank>_time_label
Labels: Forward-looking, Not Forward-looking

Intended Use
This model is designed for researchers and practitioners working on subjective text classification, particularly within financial contexts. It assesses the Temporal Classification attribute, aiding in the analysis of subjective content in financial communications.

How to Use

from transformers import pipeline, AutoTokenizer, AutoModelForSequenceClassification, AutoConfig

# Load tokenizer, model, and configuration
tokenizer = AutoTokenizer.from_pretrained("gtfintechlab/model_{bank}_time_label", do_lower_case=True, do_basic_tokenize=True)
model = AutoModelForSequenceClassification.from_pretrained("gtfintechlab/model_{bank}_time_label", num_labels=2)
config = AutoConfig.from_pretrained("gtfintechlab/model_{bank}_time_label")

# Initialize text classification pipeline
classifier = pipeline('text-classification', model=model, tokenizer=tokenizer, config=config, framework="pt")

# Classify Temporal Classification
sentences = [
  "[Sentence 1]",
  "[Sentence 2]"
]
results = classifier(sentences, batch_size=128, truncation="only_first")
print(results)

Label Interpretation

LABEL_0: Forward-looking; discusses future economic events or decisions.
LABEL_1: Not Forward-looking; discusses past or current economic events or decisions.

3. Uncertainty Estimation

Model Name Pattern: model_<bank>_certainty_label
Labels: Certain, Uncertain

How to Use

from transformers import pipeline, AutoTokenizer, AutoModelForSequenceClassification, AutoConfig

# Load tokenizer, model, and configuration
tokenizer = AutoTokenizer.from_pretrained("gtfintechlab/model_{bank}_certainty_label", do_lower_case=True, do_basic_tokenize=True)
model = AutoModelForSequenceClassification.from_pretrained("gtfintechlab/model_{bank}_certainty_label", num_labels=2)
config = AutoConfig.from_pretrained("gtfintechlab/model_{bank}_certainty_label")

# Initialize text classification pipeline
classifier = pipeline('text-classification', model=model, tokenizer=tokenizer, config=config, framework="pt")

# Classify Uncertainty Estimation
sentences = [
  "[Sentence 1]",
  "[Sentence 2]"
]
results = classifier(sentences, batch_size=128, truncation="only_first")
print(results)

Label Interpretation

LABEL_0: Certain; presents information definitively.
LABEL_1: Uncertain; presents information with speculation, possibility, or doubt.

Running the Benchmarking Scripts

LLM Benchmarking

All commands are executed from src/llm_benchmarking:

Run every bank × 3 seeds for a given prompt format:

./run_all.sh <prompt_format>

prompt_format ∈ [guide (with annotation guides), few_shot (few‑shot examples), no_guide (zero‑shot)]

Custom experiment

Edit root/src/llm_benchmarking/template.yaml (temperature, max_tokens, feature, etc.), then:

python inference.py --config template.yaml

FinMA inference

python finma_inference.py --config template.yaml

Pre-Trained Language Models Benchmarking

Aggregated‑dataset run

Run this from the root/src/plm_benchmarking;

python general_setup_run.py

Bank‑specific setup

Run this from the root/src/plm_benchmarking;

python bank_specific_setup_run.py

Results

Table 1.  F1 scores for Stance Detection in the General Setup (Hawkish / Dovish / Neutral / Irrelevant)

Standard deviations in parentheses.
Best‑performing PLM cells are in bold; best‑performing LLM cells are in italics.

Model abbreviation key – click to expand

Abbrev	Model (checkpoint)	Category	Params
MBB	`ModernBERT‑base`	PLM – Base	150 M
BB	`bert‑base‑uncased`	PLM – Base	110 M
FB	`finbert‑pretrain`	PLM – Base	110 M
RBB	`roberta‑base`	PLM – Base	125 M
MBL	`ModernBERT‑large`	PLM – Large	396 M
BL	`bert‑large`	PLM – Large	340 M
RBL	`roberta‑large`	PLM – Large	355 M
FM	`finma‑7b‑full`	LLM (Closed)	7 B
Gem	`gemini‑2.0‑flash`	LLM (Closed)	—
4o	`gpt‑4o‑2024‑08‑06`	LLM (Closed)	—
4.1	`gpt‑4.1‑2025‑04‑14`	LLM (Closed)	—
4.1M	`gpt‑4.1‑mini‑2025‑04‑14`	LLM (Closed)	—
DS	`DeepSeek‑V3‑0324`	LLM (Open)	671 B
Qwen	`Qwen2.5‑72B‑Instruct‑Turbo`	LLM (Open)	72.7 B
L3	`Llama‑3‑70b‑chat‑hf`	LLM (Open)	70 B
L4S	`Llama‑4‑Scout‑17B‑16E‑Instruct`	LLM (Open)	405 B

Central Banks' abbreviation key – click to expand

Abbrev	Full Name
FOMC	Federal Open Market Committee (USA)
PBoC	People's Bank of China
BoJ	Bank of Japan
BoE	Bank of England
SNB	Swiss National Bank
BCB	Central Bank of Brazil
RBI	Reserve Bank of India
ECB	European Central Bank
CBR	Central Bank of the Russian Federation
CBCT	Central Bank of China (Taiwan)
MAS	Monetary Authority of Singapore
BoK	Bank of Korea (South Korea)
RBA	Reserve Bank of Australia
BoI	Bank of Israel
BoC	Bank of Canada
BdeM	Bank of Mexico
NBP	Narodowy Bank Polski (Poland)
CBRT	Central Bank of Turkey
BoT	Bank of Thailand
CBE	Central Bank of Egypt
BNM	Bank Negara Malaysia
BSP	Central Bank of the Philippines
CBoC	Central Bank of Chile
BCRP	Central Reserve Bank of Peru
BanRep	Bank of the Republic (Colombia)

Bank	MBB	BB	FB	RBB	MBL	BL	RBL	Gem	4o	4.1m	4.1	DS	Qwen	FM	L3	L4S
BCB	.678 (.039)	.635 (.057)	.609 (.008)	.655 (.018)	.673 (.016)	.636 (.014)	.634 (.050)	.528 (.027)	.498 (.017)	.462 (.020)	.504 (.017)	.613 (.021)	.525 (.028)	.350 (.034)	.503 (.021)	.589 (.049)
BCRP	.788 (.021)	.781 (.008)	.764 (.009)	.779 (.028)	.798 (.013)	.801 (.024)	.821 (.015)	.675 (.004)	.628 (.008)	.634 (.035)	.665 (.004)	.666 (.010)	.503 (.062)	.301 (.031)	.641 (.014)	.620 (.043)
BNM	.626 (.029)	.629 (.017)	.653 (.023)	.601 (.012)	.630 (.041)	.644 (.013)	.640 (.009)	.409 (.006)	.443 (.027)	.430 (.007)	.409 (.025)	.475 (.013)	.333 (.026)	.160 (.033)	.567 (.025)	.435 (.005)
BSP	.741 (.015)	.697 (.020)	.707 (.028)	.749 (.010)	.724 (.049)	.698 (.025)	.784 (.017)	.424 (.042)	.420 (.069)	.451 (.039)	.514 (.076)	.534 (.029)	.380 (.015)	.245 (.027)	.584 (.042)	.500 (.035)
BanRep	.685 (.031)	.638 (.022)	.679 (.017)	.673 (.013)	.702 (.001)	.650 (.017)	.701 (.013)	.515 (.021)	.455 (.031)	.520 (.036)	.553 (.015)	.570 (.037)	.450 (.023)	.230 (.033)	.573 (.038)	.423 (.033)
BoC	.722 (.033)	.740 (.006)	.750 (.010)	.745 (.032)	.721 (.004)	.755 (.005)	.785 (.010)	.629 (.069)	.647 (.052)	.641 (.052)	.657 (.028)	.657 (.026)	.524 (.038)	.264 (.029)	.669 (.043)	.644 (.024)
BoE	.686 (.031)	.706 (.056)	.734 (.052)	.769 (.019)	.735 (.039)	.755 (.009)	.785 (.034)	.543 (.026)	.524 (.031)	.537 (.048)	.602 (.031)	.543 (.070)	.396 (.021)	.129 (.026)	.661 (.044)	.518 (.045)
BoI	.652 (.003)	.642 (.023)	.616 (.019)	.614 (.008)	.658 (.023)	.628 (.033)	.689 (.017)	.474 (.011)	.460 (.005)	.433 (.032)	.526 (.013)	.482 (.001)	.329 (.025)	.085 (.011)	.594 (.023)	.430 (.008)
BoJ	.691 (.020)	.662 (.044)	.629 (.047)	.660 (.020)	.708 (.042)	.683 (.033)	.702 (.025)	.524 (.010)	.545 (.021)	.465 (.028)	.565 (.008)	.498 (.009)	.406 (.040)	.157 (.033)	.574 (.027)	.507 (.026)
BoK	.723 (.056)	.664 (.009)	.679 (.011)	.706 (.018)	.740 (.009)	.700 (.028)	.755 (.019)	.646 (.040)	.648 (.030)	.594 (.066)	.678 (.016)	.629 (.016)	.466 (.076)	.181 (.031)	.632 (.032)	.592 (.047)
BdeM	.696 (.020)	.694 (.048)	.642 (.032)	.724 (.023)	.716 (.008)	.684 (.026)	.735 (.030)	.596 (.009)	.602 (.024)	.509 (.023)	.626 (.016)	.669 (.034)	.447 (.047)	.118 (.018)	.642 (.012)	.552 (.013)
BoT	.717 (.046)	.696 (.064)	.723 (.021)	.735 (.057)	.733 (.071)	.728 (.006)	.741 (.017)	.547 (.004)	.549 (.032)	.551 (.039)	.573 (.012)	.581 (.029)	.484 (.038)	.258 (.026)	.596 (.009)	.577 (.030)
CBCT	.641 (.049)	.637 (.031)	.635 (.029)	.667 (.015)	.616 (.035)	.678 (.032)	.688 (.024)	.451 (.044)	.474 (.044)	.474 (.031)	.485 (.026)	.522 (.022)	.388 (.049)	.180 (.037)	.556 (.015)	.475 (.051)
CBE	.773 (.027)	.783 (.026)	.790 (.004)	.810 (.016)	.822 (.015)	.788 (.029)	.836 (.014)	.629 (.037)	.672 (.036)	.581 (.045)	.648 (.036)	.636 (.007)	.352 (.056)	.142 (.014)	.702 (.021)	.594 (.024)
CBR	.763 (.031)	.754 (.017)	.750 (.020)	.798 (.032)	.811 (.023)	.779 (.022)	.835 (.011)	.759 (.015)	.749 (.027)	.693 (.026)	.701 (.049)	.794 (.028)	.573 (.035)	.146 (.022)	.772 (.029)	.665 (.013)
CBRT	.717 (.015)	.743 (.016)	.724 (.012)	.746 (.011)	.746 (.014)	.724 (.032)	.762 (.027)	.495 (.006)	.421 (.014)	.424 (.018)	.475 (.015)	.539 (.030)	.277 (.006)	.133 (.020)	.653 (.036)	.416 (.032)
CBoC	.760 (.048)	.747 (.006)	.743 (.054)	.779 (.049)	.792 (.043)	.793 (.058)	.799 (.037)	.668 (.032)	.604 (.027)	.605 (.033)	.678 (.057)	.676 (.038)	.559 (.072)	.223 (.040)	.685 (.019)	.539 (.071)
ECB	.707 (.040)	.699 (.048)	.668 (.030)	.724 (.050)	.724 (.014)	.713 (.022)	.755 (.024)	.638 (.023)	.599 (.019)	.610 (.038)	.660 (.021)	.637 (.005)	.548 (.017)	.206 (.052)	.613 (.020)	.595 (.016)
FOMC	.671 (.029)	.674 (.035)	.675 (.046)	.747 (.020)	.732 (.037)	.685 (.044)	.749 (.047)	.572 (.021)	.584 (.025)	.564 (.018)	.649 (.023)	.653 (.023)	.512 (.023)	.170 (.025)	.599 (.012)	.498 (.015)
MAS	.656 (.049)	.681 (.043)	.666 (.005)	.666 (.066)	.698 (.042)	.680 (.049)	.703 (.033)	.553 (.046)	.581 (.026)	.588 (.034)	.569 (.015)	.689 (.041)	.540 (.026)	.347 (.024)	.638 (.035)	.646 (.023)
NBP	.685 (.016)	.690 (.013)	.705 (.026)	.731 (.013)	.725 (.009)	.697 (.010)	.695 (.020)	.637 (.015)	.631 (.043)	.614 (.063)	.665 (.031)	.660 (.002)	.508 (.028)	.118 (.017)	.618 (.035)	.597 (.015)
PBoC	.791 (.033)	.763 (.008)	.742 (.046)	.787 (.026)	.813 (.014)	.793 (.007)	.786 (.017)	.492 (.046)	.559 (.037)	.531 (.026)	.535 (.033)	.592 (.037)	.379 (.017)	.128 (.018)	.613 (.033)	.446 (.024)
RBA	.685 (.023)	.681 (.019)	.682 (.027)	.672 (.015)	.707 (.029)	.695 (.024)	.741 (.028)	.531 (.049)	.478 (.079)	.483 (.058)	.553 (.074)	.537 (.055)	.358 (.049)	.133 (.020)	.614 (.034)	.495 (.057)
RBI	.604 (.037)	.649 (.035)	.653 (.041)	.628 (.034)	.633 (.032)	.655 (.039)	.668 (.044)	.489 (.025)	.519 (.041)	.509 (.026)	.495 (.027)	.542 (.016)	.431 (.008)	.231 (.058)	.581 (.030)	.557 (.032)
SNB	.692 (.017)	.685 (.028)	.653 (.016)	.685 (.009)	.725 (.018)	.698 (.030)	.713 (.020)	.635 (.003)	.601 (.015)	.640 (.037)	.652 (.016)	.643 (.024)	.554 (.029)	.252 (.008)	.612 (.014)	.607 (.051)
Average	.702 (.030)	.695 (.028)	.691 (.025)	.714 (.025)	.723 (.026)	.710 (.025)	.740 (.024)	.562 (.085)	.556 (.084)	.542 (.075)	.586 (.078)	.601 (.075)	.449 (.083)	.196 (.071)	.620 (.053)	.541 (.074)

For more results and a comprehensive leaderboard, visit the WCB website.

License

The WCB dataset is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC-BY-NC-SA 4.0) license, which allows others to share, copy, distribute, and transmit the work, as well as to adapt the work, provided that appropriate credit is given, a link to the license is provided, and any changes made are indicated.

Citation: If you use our open-source dataset or refer to our results, please cite our paper:

@article{WCBShahSukhaniPardawala,
  title={Words That Unite The World: A Unified Framework for Deciphering Global Central Bank Communications},
  author={Agam Shah, Siddhant Sukhani, Huzaifa Pardawala et. al},
  year={2025}
}

Issue Reporting

For any questions or concerns, please open a GitHub issue or email: huzaifahp7@gmail.com, ashah482@gatech.edu, siddhantsukhani5@gmail.com

Name		Name	Last commit message	Last commit date
Latest commit History 3 Commits
cleaned_data		cleaned_data
configs		configs
croissant_files		croissant_files
final_data		final_data
llm_inference_logs		llm_inference_logs
llm_inference_outputs		llm_inference_outputs
raw_data		raw_data
resources		resources
sanitized_data		sanitized_data
src		src
utils		utils
.env.sample		.env.sample
.gitattributes		.gitattributes
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
master_file_metadata.json		master_file_metadata.json
master_file_metadata.xlsx		master_file_metadata.xlsx
requirements.txt		requirements.txt

License

gtfintechlab/WorldCentralBanks

Folders and files

Latest commit

History

Repository files navigation

Words That Unite The World: A Unified Framework for Deciphering Central‑Bank Communications Globally

Authors

Paper & Website

Dataset Overview

Model Information

Annotation Details

Abstract

Repository Layout

Format of Released Data

Released Splits

Metadata

Environment Setup

Code - getting started

Bank Dictionary: Mapping Dataset Keys to Official Central Bank Names

Dataset Availability

Access the Dataset on Hugging Face

Loading the Full Corpus (380k Sentences)

Loading the Aggregated Dataset (25k Annotated Sentences)

Loading a Specific Central Bank's Dataset

Available Seeds

Croissant Files

Loading Models from Hugging Face

Using Pre-Trained Models for WCB Tasks

1. Stance Detection

2. Temporal Classification

3. Uncertainty Estimation

Running the Benchmarking Scripts

LLM Benchmarking

Run every bank × 3 seeds for a given prompt format:

Custom experiment

FinMA inference

Pre-Trained Language Models Benchmarking

Aggregated‑dataset run

Bank‑specific setup

Results

Table 1. F1 scores for Stance Detection in the General Setup (Hawkish / Dovish / Neutral / Irrelevant)

License

Citation: If you use our open-source dataset or refer to our results, please cite our paper:

Issue Reporting

About

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Contributors 3

Uh oh!

Languages

Words That Unite The World: A Unified Framework for Deciphering Central‑Bank Communications Globally

Table 1.  F1 scores for Stance Detection in the General Setup (Hawkish / Dovish / Neutral / Irrelevant)

Packages