TakueGhost/NhauFinance-corpus

NhauFinance: African Central Bank NLP Corpus and Benchmark

The first domain-adapted language model and annotated benchmark dataset targeting African central bank discourse.

---

Dataset Summary

NhauFinance is a corpus of 1,322 documents (40 million tokens) from five African monetary authorities, accompanied by 1,876 expert-annotated examples across four downstream NLP tasks. Note: the HuggingFace size_categories tag (1K<n<10K) refers to the number of labeled examples (2,800 across all tasks), not the full corpus document count.

The corpus was constructed to address a concrete and consequential gap: African central bank communications are analytically inaccessible to standard financial NLP tools. ProsusAI/FinBERT scores zero F1 on CURRENCY entity recognition for African monetary terminology (ZiG, RTGS dollar, cedi), confirming a vocabulary-level domain gap that this dataset is designed to close.

The accompanying model, NhauFinance-v2, achieves a 6x data efficiency advantage on financial distress detection: reaching 0.62 macro F1 with just 50 labeled examples, compared to ~300 examples required by FinBERT.

---

Institutional Coverage

Institution	Abbreviation	Documents	Approx. Tokens	Primary Document Types
South African Reserve Bank	SARB	572	~12.1M	MPC Statements, MPRs, QPRs
Central Bank of Nigeria	CBN	362	~10.4M	MPC Communiques, Annual Reports, FSRs
Central Bank of Kenya	CBK	298	~9.8M	MPC Reports, Monetary Policy Reviews
Bank of Ghana	BOG	52	~4.7M	Monetary Policy Reports, FSRs
Reserve Bank of Zimbabwe	RBZ	38	~3.0M	Monetary Policy Statements, Exchange Rate Notices
Total		1,322	~40M

---

Benchmark Tasks

Task 1: Monetary Policy Sentiment

• Classes: HAWKISH, DOVISH, NEUTRAL
• Examples: ~440 usable (R1+R2+R3, OUT_OF_DOMAIN excluded)
• Baseline (FinBERT macro F1): 0.5623 +/- 0.0216
• NhauFinance-v2 (macro F1): 0.4608 +/- 0.0291

Task 2: Financial Distress Detection

• Classes: STRESS, STABLE
• Examples: ~545 usable
• Baseline (FinBERT macro F1): 0.6790 +/- 0.0497
• NhauFinance-v2 (macro F1): 0.7209 +/- 0.0506 (Delta = +4.20 pts)
• Key finding: NhauFinance-v2 achieves 0.62 macro F1 with 50 labeled examples; FinBERT requires ~300

Task 3: Policy Event Classification

• Classes: RATE_DECISION, INTERVENTION, CURRENCY_ACTION, REGULATORY_CHANGE, NO_EVENT
• Examples: ~515 usable
• Baseline (FinBERT macro F1): 0.4713 +/- 0.1018
• NhauFinance-v2 (macro F1): 0.5491 +/- 0.1237 (Delta = +7.78 pts)
• Note: CURRENCY_ACTION (n=17) and INTERVENTION (n=10) are low-support classes reported with full disclosure

Task 4: Named Entity Recognition

• Classes: CENTRAL_BANK, CURRENCY, POLICY_RATE, ECONOMIC_INDICATOR, INSTITUTION
• Examples: ~891
• Baseline (FinBERT macro F1): 0.3852 +/- 0.0480
• FinBERT CURRENCY F1: 0.0000 (domain gap confirmed)
• NhauFinance-v2 (macro F1): 0.3514 +/- 0.0336 (51% lower cross-seed variance)

---

Repository Structure

---

Labeled Data Schema

Tasks 1, 2, 3 (Classification)

Column	Description
text	Snippet text (~150 words, keyword-anchored for event tasks)
label	Task-specific class label
bank	Source institution (CBN, SARB, CBK, RBZ, BOG)
filename	Source document filename
subfolder	Document type subfolder
confidence	Panel confidence level (HIGH, MEDIUM, LOW)
round	Annotation round (1, 2, 3)
label_source	panel or fallback_error
hanke_veto_applied	Boolean — RBZ/CBN Hanke-Chelwa dual validation applied
institutional_survival_flag	Boolean — Expert 5C flag for performance vs genuine signal

Task 4 (NER)

Column	Description
text	Snippet text
entities	JSON list of entity dicts with text, type, optional start/end
bank	Source institution
round	Annotation round

---

Annotation Methodology

Annotations were produced via an LLM-assisted expert simulation protocol modelling a 17-member virtual expert panel spanning:

• Category 1: Emerging Market Practitioners (EM asset management perspective)
• Category 2: Quantitative Researchers (systematic signal extraction standards)
• Category 3: Academic Reviewers (JFDS, JIE, African Finance Journal)
• Category 4: NLP Technical Reviewers
• Category 5: Domain Validators — CBN MPC Economist (5A), RBZ authority (5B), INSTITUTIONAL_SURVIVAL flag (5C)

Quality controls: Fallback error rate monitored (<10% threshold); Hanke-Chelwa dual validation for RBZ and CBN 2016-2023; distribution checks after each 50-snippet batch; OUT_OF_DOMAIN exclusion for evaluation.

Independent human validation of a stratified annotation subset is in progress and will be reported in the accompanying working paper.

---

Key Finding: Institutional Survival Framing

Analysis of annotation patterns reveals that 96.2% of African CB policy event announcements are embedded in institutional survival framing — a structural feature of African sovereign communication absent from developed-market NLP frameworks. This framing pattern, in which policy announcements are organised around assertions of institutional continuity or sovereign creditworthiness, has direct implications for how EM sovereign risk signals should be extracted from African CB text.

---

Pre-trained Model

The companion model NhauFinance-v2 will be released at: https://huggingface.co/TakueGhost/NhauFinance-v2

• Initialised from ProsusAI/FinBERT
• Pre-trained on 40M tokens of NhauFinance corpus
• 12,230 training steps
• Validation perplexity: 3.84
• Validation loss: 1.345

---

Citation

If you use this dataset or the NhauFinance-v2 model in your research, please cite:

@misc{nhaufinance2026,
  title        = {NhauFinance: A Domain-Adapted Language Model and Benchmark
                  for African Central Bank Natural Language Processing},
  author       = {Chirindo, Takudzwa},
  year         = {2026},
  howpublished = {Working Paper, NhauFinance Research Group},
  note         = {HuggingFace: TakueGhost/NhauFinance-corpus.
                  SSRN: [update with URL after posting tonight]},
  url          = {https://huggingface.co/datasets/TakueGhost/NhauFinance-corpus}
}

---

License

This dataset is released under Creative Commons Attribution 4.0 International (CC BY 4.0).

You are free to share and adapt the material for any purpose, provided appropriate credit is given.

---

Applications

This dataset is designed to support:

• Sovereign risk monitoring: Identifying financial stress signals in African CB communications ahead of credit rating actions
• EM monetary policy analysis: Tracking hawkish/dovish shifts across five African monetary regimes
• Policy event extraction: Automated detection of rate decisions, FX interventions, and currency actions
• Central bank NLP research: Establishing reproducible benchmarks for the underserved African CB NLP domain

---

Contact

For questions, collaboration, or to contribute additional African CB document coverage:

• HuggingFace: @TakueGhost
• Repository discussions: TakueGhost/NhauFinance-corpus/discussions