bilalezafar/centralbank-bert

Purpose and Execution Model

This is a notebook-driven NLP research pipeline for analyzing CBDC discourse in central-bank text. There is no application service or package entry point; notebooks are the executable entry points for model training, inference, evaluation, and publication-result generation.

1. Domain Encoder Foundation

Domain adaptation starts with masked-language-model training over BIS speeches, producing the CentralBank-BERT encoder used by downstream tasks.

• 1-Central Bank BERT (MLM)/CentralBank-BERT.ipynb

The resulting encoder is externally published on Hugging Face as bilalezafar/CentralBank-BERT; training data and model artifacts are not represented as a reusable local code module.

2. CBDC Detection Gate

The binary CBDC-vs-non-CBDC classifier is the gating stage for broad corpora: text extraction and keyword-assisted data preparation feed fine-tuning, then accepted CBDC material can receive finer labels.

• 2-CBDC-BERT/CBDC speeches data extraction for BERT.ipynb

• 2-CBDC-BERT/CBDC-BERT model training and evaluation.ipynb

• 2-CBDC-BERT/bert_training_data.csv

cbdc_keywords.csv is a material extraction boundary. The trained classifier is served externally as bilalezafar/CBDC-BERT.

3. Independent CBDC Attribute Classifiers

Four separately trained sentence-level classifiers consume CBDC text and predict complementary dimensions. Each notebook owns preprocessing, tokenization, fine-tuning, and evaluation, backed by its labeled CSV.

• Stance—Pro, Wait-and-See, Anti: 3-CBDC-Stance/cbdc-stance.ipynb, 3-CBDC-Stance/stance_sentences.csv

• Sentiment—Positive, Neutral, Negative: 4-CBDC-Sentiment/cbdc-sentiment.ipynb, 4-CBDC-Sentiment/cbdc_sentiment_training.csv

• Type—Retail, Wholesale, General: 5-CBDC-Type/cbdc-type.ipynb, 5-CBDC-Type/cbdc_type_training.csv

• Discourse—Feature, Process, Risk-Benefit: 6-CBDC-Discourse/cbdc-discourse.ipynb, 6-CBDC-Discourse/cbdc_classification_training.csv

These models are released separately through Hugging Face, rather than versioned as local weights.

4. Central-Bank NER Extension

NER is a parallel task built on the domain-adapted encoder, intended to identify entities in central-banking discourse rather than enrich the CBDC classifier output directly.

• 7-CentralBank-NER/CentralBank-NER.ipynb

• 7-CentralBank-NER/Access link to BIS Speeches NER dataset.docx

Its dataset is externally accessed, making dataset availability a reproducibility boundary.

5. Integrated Corpus Inference

The primary end-to-end analysis entry point reads a finalized CBDC corpus, invokes the classifier suite, and writes row-level predictions.

• 8-CBDC-Analysis and Results/CBDC Analysis and Resutls.ipynb

• 8-CBDC-Analysis and Results/cbdc-dataset-final.csv

• 8-CBDC-Analysis and Results/cbdc-dataset-predictions.csv

Control flow is corpus input → CBDC gate/attribute predictions → merged prediction dataset → aggregate analysis. The prediction CSV is the central handoff between ML inference and research reporting.

6. Results, Robustness, and Benchmarking

Post-inference outputs are materialized as CSV tables for label distributions, time series, crosstabs, confidence review, and exemplars.

• 8-CBDC-Analysis and Results/prediction_results/

• 8-CBDC-Analysis and Results/descriptive_results/

• 8-CBDC-Analysis and Results/compare_out/Comparision - Robustness.ipynb

A matched encoder benchmark compares BERT-base-uncased against CentralBank-BERT across the CBDC tasks:

• 8-CBDC-Analysis and Results/CBDC classifiers benchmark with BERT-base.ipynb

Runtime and External Dependencies

The runtime is Python/Jupyter with transformer-based BERT fine-tuning and inference. Hugging Face is the key external model distribution service. BIS speeches underpin encoder training and NER data access; the repository retains local labeled datasets and generated analytical outputs, but not a deployable API, orchestration system, or centralized pipeline library.