README.md

---
license: cc-by-sa-4.0
task_categories:
  - question-answering
  - text-retrieval
language:
  - en
size_categories:
  - 1M<n<10M
pretty_name: OPERA Retrieval Corpus
tags:
  - multi-hop-qa
  - dense-retrieval
  - bge-m3
  - faiss
  - opera
---

# OPERA Retrieval Corpus

Prebuilt dense retrieval corpus used by the OPERA inference pipeline
(*OPERA: A Reinforcement Learning–Enhanced Orchestrated Planner-Executor
Architecture for Reasoning-Oriented Multi-Hop Retrieval*).

This release contains two files:

| File | Size | Purpose |
|---|---|---|
| `opera_corpus.index` | ~7.3 GB | FAISS `IndexFlatIP` over BGE-M3 dense embeddings |
| `opera_corpus.meta`  | ~394 MB | Pickle list of `{id, title, content}` aligned 1:1 with the FAISS rows |

- Total documents: **1,780,294**
- Embedding model: **`BAAI/bge-m3`** (1024-dim, cosine via inner-product on L2-normalised vectors)
- Index type: **FAISS `IndexFlatIP`** (exact search, no quantization)

## Source

The corpus is composed of Wikipedia paragraphs aggregated from three
public multi-hop QA benchmarks evaluated in the OPERA paper:

- **HotpotQA** (Yang et al., 2018)
- **2WikiMultiHopQA** (Ho et al., 2020)
- **MuSiQue** (Trivedi et al., 2022)

All text in the corpus comes from the Wikipedia paragraphs distributed
with these benchmarks. No external web content was added, and no
benchmark labels (supporting-paragraph flags, sub-question
decompositions, gold answers) are included — fetch those from the
upstream datasets if you need them.

## Schema

Each record in `opera_corpus.meta` is a Python dict with exactly three
keys:

```python
{
    "id":      "doc_0",                    # sequential, stable across loads
    "title":   "The Big Short (film)",     # source-document title
    "content": "The Big Short is a 2015 American ...",
}
```

Row `i` of `opera_corpus.meta` corresponds to row `i` of
`opera_corpus.index`. The row order has been shuffled with a fixed
seed; the IDs `doc_0 … doc_1780293` are simple sequential identifiers
and carry no information about which benchmark a paragraph came from.

Content lengths vary across records (paragraph-level and shorter
fragments coexist in the index). If your downstream use case prefers a
single granularity, filter by `len(record["content"])`.

## Loading

```python
import pickle, faiss

index = faiss.read_index("opera_corpus.index")
with open("opera_corpus.meta", "rb") as f:
    meta = pickle.load(f)

assert index.ntotal == len(meta) == 1780294
print(meta[0])
```

See `load_example.py` for an end-to-end retrieval example with BGE-M3.

## Reproducing OPERA inference

The OPERA inference pipeline (Plan Agent → BGE-M3 retriever →
Analysis-Answer Agent → Rewrite Agent → final synthesis) is released
separately at the OPERA repository. Point its `--index-path` and
`--metadata-path` flags at the two files in this release, or serve them
behind the persistent retriever HTTP server shipped with the pipeline.

**Evaluation splits.** The reproducibility numbers below are produced
on random 500-question subsamples of the official **dev splits** of
HotpotQA, 2WikiMultiHopQA, and MuSiQue, drawn with fixed seeds (42 and
53). For MuSiQue we stratify the 500-question sample across hop counts
(350 × 2-hop + 100 × 3-hop + 50 × 4-hop) to keep the difficulty
distribution close to the full dev set. Two independent seeds are used
so variance across runs can be reported.

On these splits, the inference-only pipeline against this corpus
reproduces the **OPERA (CoT)** row of Table 1 of the paper, i.e. the
OPERA architecture *without* MAPGRPO training:

| Split  | HotpotQA EM | 2Wiki EM | MuSiQue EM |
|---|---|---|---|
| seed42 | 41.8 | 44.0 | 24.0 |
| seed53 | 42.2 | 42.2 | 20.8 |
| paper "OPERA (CoT)" | 44.9 | 42.3 | 21.2 |

Headline numbers in Table 1 of the paper (the OPERA-MAPGRPO row)
require MAPGRPO training of the three agents; the training code is not
part of this release.

## Citation

```bibtex
@article{Liu2026OPERA,
  author  = {Liu, Yu and Liu, Yanbing and Yuan, Fangfang and Cao, Cong and Sun, Youbang and Peng, Kun and Chen, WeiZhuo and Li, Jianjun and Ma, Zhiyuan},
  title   = {{OPERA}: A Reinforcement Learning--Enhanced Orchestrated Planner-Executor Architecture for Reasoning-Oriented Multi-Hop Retrieval},
  journal = {arXiv preprint arXiv:2508.16438},
  year    = {2025}
}
```

## License

The retrieval index and metadata are released for research use only.
Underlying paragraph text comes from HotpotQA, 2WikiMultiHopQA and
MuSiQue and remains subject to those datasets' original licenses, each
of which derives from Wikipedia under CC-BY-SA 4.0. When publishing
results obtained with this corpus, please cite the OPERA paper as well
as the original benchmark papers:

- Yang et al., *HotpotQA*, EMNLP 2018.
- Ho et al., *Constructing a multi-hop QA dataset for comprehensive
  evaluation of reasoning steps*, COLING 2020.
- Trivedi et al., *MuSiQue: Multihop Questions via Single-hop Question
  Composition*, TACL 2022.
- Chen et al., *BGE M3-Embedding*, 2024.

## Contact

For questions about the corpus or the OPERA pipeline, please open a
discussion on this dataset's HuggingFace page or an issue on the OPERA
GitHub repository.