README.md

---
license: cc-by-4.0
task_categories:
  - question-answering
language:
  - en
tags:
  - knowledge-graph
  - biomedical
  - sparql
  - compositional-qa
  - multi-hop
pretty_name: KORA Benchmark
size_categories:
  - 1K<n<10K
configs:
- config_name: default
  data_files:
  - split: full
    path: benchmark/benchmark_clean.csv
- config_name: splits
  data_files:
  - split: train
    path: benchmark/train.csv
  - split: validation
    path: benchmark/val.csv
  - split: test
    path: benchmark/test.csv
---

# KORA Benchmark

Resources for reproducing **KORA: Adaptive Multi-Agent Orchestrated Retrieval over Knowledge Graphs** — including the BioCQ benchmark dataset, entity resolution indexes, and the combined biomedical knowledge graph.

## Repository Contents

| Path | Description |
|------|-------------|
| `benchmark/` | BioCQ question splits (train / val / test / full) |
| `indexes/scispacy*/` | Pre-built SciSpaCy entity resolution indexes (~1 GB) |
| `indexes/ark_bm25/` | Pre-built ARK BM25 retrieval indexes (~406 MB) |
| `kg/kg.nt` | Combined RDF knowledge graph (PrimeKG + OptimusKG + AfrOMedKG) in N-Triples format |
| `kg/virtuoso_clean.ini` | Virtuoso SPARQL server config template |

## BioCQ Dataset

**8,578 compositional biomedical questions** across five reasoning types:

| Question Type | Questions |
|---|---:|
| Multi-hop Traversal | 2,094 |
| Aggregation & Counting | 1,991 |
| Constrained Retrieval | 1,784 |
| Differential Diagnosis | 1,437 |
| Knowledge Gap | 1,272 |
| **Total** | **8,578** |

Questions are grounded in three biomedical knowledge graphs: **PrimeKG**, **OptimusKG**, and **AfrOMedKG**.

## Usage

### Load the benchmark

```python
import pandas as pd

train = pd.read_csv("benchmark/train.csv")
val   = pd.read_csv("benchmark/val.csv")
test  = pd.read_csv("benchmark/test.csv")
```

Or via the HuggingFace datasets library:

```python
from datasets import load_dataset

# Full dataset (8,578 questions)
ds = load_dataset("anonymous-kgqa/KORA-Benchmark")["full"]

# Pre-defined train / val / test splits
splits = load_dataset("anonymous-kgqa/KORA-Benchmark", "splits")
train, val, test = splits["train"], splits["validation"], splits["test"]
```

### Download SciSpaCy entity resolution indexes

```bash
hf download anonymous-kgqa/KORA-Benchmark --repo-type=dataset \
  --include "indexes/scispacy*" --local-dir kora/kg/index_data
```

### Download ARK BM25 indexes

```bash
hf download anonymous-kgqa/KORA-Benchmark --repo-type=dataset \
  --include "indexes/ark_bm25/*" --local-dir baselines/ark/temp && \
mv baselines/ark/temp/indexes/ark_bm25/*.pkl baselines/ark/temp/ && \
rm -rf baselines/ark/temp/indexes
```

### Download the knowledge graph

```bash
hf download anonymous-kgqa/KORA-Benchmark --repo-type=dataset \
  --include "kg/*" --local-dir data/kg
```

Load `kg/kg.nt` into Virtuoso using the provided `kg/virtuoso_clean.ini` template. See the KORA repository for full setup instructions.


## License

[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)