README.md

---
license: apache-2.0
task_categories:
- text-classification
language:
- en
- sw
tags:
- esg
- esg-disclosure
- sustainability
- finance
- nairobi-securities-exchange
- kenya
size_categories:
- 1K<n<10K
---
# KenyaESG: Sentence-level ESG Disclosure Classification Dataset (NSE)

Sentence-level dataset for three independent binary ESG classification tasks (Environmental,
Social, Governance) built from corporate reports of firms listed on the Nairobi Securities
Exchange (NSE), 2010–2024, combined with reference sentences from Schimanski et al. (2024).
Companion data to the fine-tuned classifiers `josephagossa/KenyaESG-RoBERTa-{env,soc,gov}`.

## Structure

Three pillars, each with a training set and a disjoint human-annotated evaluation set, provided
as Excel files (one per pillar):

| Pillar | Training file | Evaluation file | train | eval |
|---|---|---|---|---|
| Environmental | KenyaESG_train_environmental.xlsx | KenyaESG_evaluation_environmental.xlsx | 3,900 | 100 |
| Social | KenyaESG_train_social.xlsx | KenyaESG_evaluation_social.xlsx | 3,900 | 100 |
| Governance | KenyaESG_train_governance.xlsx | KenyaESG_evaluation_governance.xlsx | 3,900 | 100 |

The training and evaluation sentences are disjoint (the 100 evaluation sentences per pillar were
removed from training to prevent leakage). Together they reconstruct the full 4,000-sentence pool
per pillar. Because the three pillars are modelled as independent binary tasks, a sentence may be
positive on more than one pillar.

You can read the files directly with pandas:

```python
import pandas as pd
train = pd.read_excel("KenyaESG_train_environmental.xlsx")   # or social / governance
test  = pd.read_excel("KenyaESG_evaluation_environmental.xlsx")
```

## Fields

Training files (`KenyaESG_train_<pillar>.xlsx`): a row index, `text`, `<pillar>` (the 0/1 label,
named `env`, `soc`, or `gov`), and `source`.

Evaluation files (`KenyaESG_evaluation_<pillar>.xlsx`): `id`, `text`, `<pillar>_A`, and
`<pillar>_B` (the two independent annotators' binary judgments, e.g. `env_A`, `env_B`).
Inter-annotator agreement (Cohen's κ): 0.94 (environmental), 0.86 (social), 0.88 (governance).
Both annotations are released rather than a single adjudicated label, so users can recompute
agreement and adjudicate themselves; the three evaluation samples are drawn independently per
pillar (the sentences differ across pillars).

## Companion models (performance vs. human benchmark)

| Model | Pillar | F1 | DOI |
|---|---|---|---|
| KenyaESG-RoBERTa-env | Environmental | 0.917 | 10.57967/hf/9126 |
| KenyaESG-RoBERTa-soc | Social | 0.882 | 10.57967/hf/9127 |
| KenyaESG-RoBERTa-gov | Governance | 0.916 | 10.57967/hf/9128 |

## Sources

The `source` field distinguishes two origins. **Kenya:** sentences from annual, integrated, and
sustainability reports of NSE-listed firms (2010–2024); training labels were assigned by a
keyword-based filter and refined by a single-reviewer pass (not full manual annotation).
**Schimanski et al. (2024):** reference sentences from "Bridging the gap in ESG measurement"
(Finance Research Letters 61, 104979), available under Apache-2.0.

## Privacy and anonymization

Company and personal names are masked in the `text` field (`[COMPANY]`, `[PERSON]`) in line with
the anonymisation principles of the Kenya Data Protection Act, 2019 (No. 24 of 2019) and
Regulation (EU) 2016/679 (GDPR). Masking of personal names follows these principles; company
names (from public filings) are masked as a precaution. Masking is semi-automatic and not
guaranteed exhaustive; a small number of false-positive masks on common words (e.g. "equity",
"equality", "Standard") were subsequently restored — see `KenyaESG_correction_log.txt`. The
dataset must not be used to attempt re-identification of individuals or firms.

## Intended use and limitations

Intended for training and evaluating sentence-level ESG text classifiers. The training labels for
the Kenya portion are keyword-derived and single-reviewer-checked; the evaluation files are the
appropriate benchmark for honest performance estimates. The dataset captures disclosure intensity,
not disclosure quality; sample sizes are modest. The corpus is predominantly English but retains a
small number of Swahili and OCR-degraded sentences, reflecting the original NSE reports.

## License and citation

© 2026 Joseph Agossa. The dataset compilation, structure, and annotations are licensed under
Apache-2.0; the underlying verbatim report text remains under the copyright of the respective
issuing companies and is included for research purposes under a fair-dealing rationale, with
sources documented.

This dataset has two persistent identifiers, which may be used interchangeably: the Hugging Face
DOI 10.57967/hf/9193 and the Zenodo concept DOI 10.5281/zenodo.20608236 (resolving to the latest
archived version).

```bibtex
@misc{agossa2026kenyaesg,
  author    = {Agossa, Joseph},
  title     = {KenyaESG: A Sentence-level ESG Disclosure Classification Dataset
               for Nairobi Securities Exchange Corporate Reports},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.20608236},
  note      = {Also available on the Hugging Face Hub, DOI 10.57967/hf/9193}
}
```

**Working paper:** Agossa, J. (2026). *Pricing the Cost of Compliance: Equity Reactions to Mandatory ESG Disclosure in a Frontier Market.* SSRN: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6966682

```bibtex
@unpublished{agossa2026compliance,
  author = {Agossa, Joseph},
  title  = {Pricing the Cost of Compliance: Equity Reactions to
            Mandatory ESG Disclosure in a Frontier Market},
  year   = {2026},
  note   = {Working paper, SSRN 6966682},
  url    = {https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6966682}
}
```

Schema adapted from Schimanski, T., Reding, A., Reding, N., Bingler, J., Kraus, M., & Leippold, M.
(2024). Bridging the gap in ESG measurement. *Finance Research Letters*, 61, 104979.