README.md

---
license: cc-by-4.0
language: [en]
tags:
- hydrology
- flood-forecasting
- retrieval-augmented-generation
- knowledge-base
- causal-reasoning
- northumbria
- defra
- environment-agency
pretty_name: HydroCause Knowledge Corpus
size_categories: [1K<n<10K]
task_categories: [text-generation, question-answering, feature-extraction]
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
---

# HydroCause Knowledge Corpus

The domain knowledge base released alongside the HydroCause-RAG framework. It is the "raw knowledge source" used by every downstream component:

- The retrieval layer (Part 2-3) grounds numerical claims against it.
- The P7 audit gate (Part 3 §3.4, SGS Audit) scores semantic reasoning against it.
- The teacher-LLM synthesis (Part 5) drafts Q&A pairs from its passages.
- The domain-tuned model HydroCause-3B-DK is fine-tuned on those Q&A pairs.

## Composition

| Sub-corpus | Passages | Source |
|---|---:|---|
| policy    |   632 | Northumberland LFRMS, DEFRA FCERM, EA guidance, Northumbria RBD FRMP |
| hydrology | 4,147 | Open-access papers (HESS/ESSD, CC BY 4.0) + curated seed summaries |
| metrics   |    12 | NSE/KGE/SMAPE/MAE/RMSE/MBE/POD/FAR/CSI/F1/CFS/SGS |
| registry  |    24 | HydroCause-RAG numerical truth registry (six catchments) with TimeSHAP-derived tau |
| **total** | **4,815** | 90/10 stratified train / validation split |

## Intended uses

This corpus supports four categories of downstream tasks:

### 1. Operational hydrological forecasting
- Grounding retrieval-augmented systems in verifiable flood-warning doctrine (WMO-1072, WMO IFM Tool 19).
- Answering operator questions about lead time, threshold selection, ensemble spread, false-alarm tradeoffs.
- Providing decision-support context for issued warnings.

### 2. Comparison of historical events to predicted
- Hindcasting methodology and extreme-event extrapolation (Frame et al. 2022).
- Return-period statistics, non-stationarity considerations, DEFRA climate-change allowances.
- Historic UK flood chronology and post-event attribution.

### 3. Informed decision-making
- Cost-loss analysis and warning-threshold calibration for a defined user community.
- Statistical hydrology for return-level estimation (Helsel 2020).
- UK Flood Forecasting Centre and Environment Agency operational practice.

### 4. General hydrological education
- Catchment processes, unit hydrograph theory, flood routing, recession analysis.
- Streamflow measurement (Rantz WSP 2175 vol 1) and discharge computation (Rantz vol 2, USGS TM 3-A8).
- Machine-learning-in-hydrology foundations (Kratzert 2018/2024, Gauch 2021, Lees 2021, Coxon 2020).
- HydroLLM-Benchmark-aligned Q&A synthesis for domain-adapted language models.

## Fields

Each row has:

- `passage_id` (str)   — SHA-1 fingerprint
- `source_label` (str) — machine-readable source identifier
- `source_url` (str)   — public URL if scraped, else `notebook_bundled` / `hydrocause_registry`
- `chunk_index` (int)  — index within the source document
- `subcorpus` (str)    — `policy` / `hydrology` / `metrics` / `registry`
- `text` (str)         — the passage
- `n_words` (int)      — word count
- `split` (str)        — `train` or `validation`

## Loading

```python
from datasets import load_dataset
ds = load_dataset("DamiOresotu/hydrocause-knowledge-corpus")
print(ds)
```

## Sources and licences

### Policy sub-corpus (UK Open Government Licence v3.0)
- **Northumberland County Council**: Local Flood Risk Management Strategy — Action Plan.
- **DEFRA**: Flood and Coastal Erosion Risk Management Policy Statement (2020); National FCERM Strategy for England.
- **DEFRA / Environment Agency**: River Basin Management Planning Ministerial Guidance.
- **Environment Agency**: Northumbria River Basin District Flood Risk Management Plan 2021-2027.

### Hydrology sub-corpus — peer-reviewed research (Copernicus, CC BY 4.0)
- Kratzert, F., Klotz, D., Brenner, C., Schulz, K., Herrnegger, M. (2018). *Rainfall-runoff modelling using Long Short-Term Memory (LSTM) networks.* Hydrology and Earth System Sciences 22:6005-6022. DOI: 10.5194/hess-22-6005-2018.
- Kratzert, F., Gauch, M., Klotz, D., Nearing, G. (2024). *HESS Opinions: Never train a Long Short-Term Memory (LSTM) network on a single basin.* HESS 28:4187. DOI: 10.5194/hess-28-4187-2024.
- Gauch, M., Kratzert, F., Klotz, D., Nearing, G., Lin, J., Hochreiter, S. (2021). *Rainfall-runoff prediction at multiple timescales with a single Long Short-Term Memory network.* HESS 25:2045-2062. DOI: 10.5194/hess-25-2045-2021.
- Coxon, G., Addor, N., Bloomfield, J.P., Freer, J., Fry, M., Hannaford, J., et al. (2020). *CAMELS-GB: hydrometeorological time series and landscape attributes for 671 catchments in Great Britain.* Earth System Science Data 12:2459-2483. DOI: 10.5194/essd-12-2459-2020.
- Lees, T., Buechel, M., Anderson, B., Slater, L., Reece, S., Coxon, G., Dadson, S.J. (2021). *Benchmarking data-driven rainfall-runoff models in Great Britain: a comparison of long short-term memory (LSTM)-based models with four lumped conceptual models.* HESS 25:5517-5534. DOI: 10.5194/hess-25-5517-2021.

### Hydrology sub-corpus — educational (public domain / CC BY 4.0)
- Kizilkaya, D., Sajja, R., Sermet, Y., Demir, I. (2025). *Towards HydroLLM: A Benchmark Dataset for Hydrology-Specific Knowledge Assessment for Large Language Models.* Environmental Data Science (Cambridge). DOI: 10.1017/eds.2025.10006. CC BY 4.0.
- Helsel, D.R., Hirsch, R.M., Ryberg, K.R., Archfield, S.A., Gilroy, E.J. (2020). *Statistical Methods in Water Resources.* U.S. Geological Survey Techniques and Methods 4-A3, 458 p. DOI: 10.3133/tm4a3. U.S. Government public domain (17 USC §105).
- Turnipseed, D.P., Sauer, V.B. (2010). *Discharge Measurements at Gaging Stations.* USGS Techniques and Methods 3-A8, 87 p. Public domain.
- Corson-Dosch, H.R. et al. (2023). *The Water Cycle.* USGS General Information Product 221. DOI: 10.3133/gip221. Public domain.
- Rantz, S.E. et al. (1982). *Measurement and Computation of Streamflow, Volume 1: Measurement of Stage and Discharge.* USGS Water-Supply Paper 2175, 284 p. DOI: 10.3133/wsp2175. Public domain.
- Rantz, S.E. et al. (1982). *Measurement and Computation of Streamflow, Volume 2: Computation of Discharge.* USGS Water-Supply Paper 2175, 347 p. DOI: 10.3133/wsp2175. Public domain.

### Hydrology sub-corpus — operational forecasting doctrine (WMO)
- World Meteorological Organization (2011). *Manual on Flood Forecasting and Warning.* WMO-No. 1072, 97 p. ISBN 978-92-63-11072-5.
- WMO / Global Water Partnership Associated Programme on Flood Management (2013). *Flood Forecasting and Early Warning.* Integrated Flood Management Tools Series No. 19.

### Hydrology sub-corpus — extreme-event research
- Frame, J.M., Kratzert, F., Klotz, D., Gauch, M., Shalev, G., Gilon, O., Qualls, L.M., Gupta, H.V., Nearing, G.S. (2022). *Deep learning rainfall-runoff predictions of extreme events.* HESS 26:3377-3392. DOI: 10.5194/hess-26-3377-2022.

### Metric definitions and seed summaries
- Metric definitions (NSE, KGE, SMAPE, MAE, RMSE, MBE, POD, FAR, CSI, F1, CFS, SGS) and concept seed summaries are original text composed for this release.

### Registry-anchored passages
- Derived from the HydroCause-RAG numerical truth registry produced by Parts 1.7-1.9 of the source notebook. Six Northumberland catchments (Acomb Gardenhouse, Acomb Millersfield, Haltwhistle, Riding Mill, Stocksfield, Morpeth) with per-catchment CFS, forecast metrics, and TimeSHAP-derived tau attributions.

The corpus (this compilation) is released under **CC BY 4.0**. Original source documents retain their original licences.

### Methodological note
The hydrology sub-corpus deliberately combines peer-reviewed research articles with educational textbook-style material and the HydroLLM-Benchmark reference paper. This mirrors the methodology of Kizilkaya et al. (2025), who built HydroLLM-Benchmark from *research articles and hydrology textbooks*, ensuring that downstream Q&A synthesis (Part 5) and retrieval (Parts 2-3) can be defended against reviewer challenges about educational versus research grounding.

## Citation

```bibtex
@article{oresotu2026hydrocauserag,
  title  = {HydroCause-RAG: A Causally Audited Retrieval-Augmented
            Framework for Explainable Time-Series Forecasting with
            Foundation-Model Portability and Local-Model Distillation},
  author = {Oresotu, Olanrewaju and Li, Yulei and Woo, Wai Lok},
  journal= {Knowledge-Based Systems},
  year   = {2026}
}
```