---
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
dataset_info:
features:
- name: query_id
dtype: int64
- name: query
dtype: string
- name: answer
dtype: string
- name: gold_docs
list:
- name: position
dtype: int64
- name: text
dtype: string
- name: url
dtype: string
- name: pass_rate
dtype: float64
splits:
- name: train
num_examples: 6102
license: mit
---
## OpenResearcher Gold Documents
This dataset contains the **gold documents** used for the "Gold Document Retrieval via Online Bootstrapping" step described in Section 3.2 of the [OpenResearcher paper](https://arxiv.org/abs/2603.20278). Gold documents are documents that collectively contain sufficient evidence to derive the ground-truth answer for a given question.
For **6,102** questions sourced from [MiroVerse](https://huggingface.co/datasets/miromind-ai/MiroVerse-v0.1), we constructed a search query by concatenating the question and reference answer, retrieved web content via the Serper API, and cleaned/deduplicated the results to obtain **~10K gold documents** (1–3 gold documents per question, averaging 2.86).
These gold documents are merged with 15M FineWeb documents (as distractors) to build the offline search corpus, [OpenResearcher-Corpus](https://huggingface.co/datasets/OpenResearcher/OpenResearcher-Corpus), which is used as a self-hosted, API-free search engine when synthesizing the deep-research trajectories released as [OpenResearcher-Dataset](https://huggingface.co/datasets/OpenResearcher/OpenResearcher-Dataset). This bootstrapping step is essential: removing it causes gold-document hit rate to drop from 29.54% to 1.73%, trajectory accuracy to drop from 56.86% to 43.81%, and downstream BrowseComp-Plus accuracy to collapse from 54.81% to 6.35% (see RQ2 in the paper).
## Format
Each row in the dataset contains the following fields:
- **query_id** (int64): A unique identifier for each question.
- **query** (string): The original question, sourced from [MiroVerse](https://huggingface.co/datasets/miromind-ai/MiroVerse-v0.1).
- **answer** (string): The reference answer used to construct the retrieval query.
- **gold_docs** (list): The gold documents retrieved and cleaned for this question. Each entry contains:
- **position** (int64): Rank position among the retrieved gold documents.
- **text** (string): The full text content of the gold document.
- **url** (string): The source URL where the document was retrieved from.
- **pass_rate** (float64): The pass rate observed for this question during trajectory synthesis with GPT-OSS-120B.
## How to use this dataset?
```python
from datasets import load_dataset
ds = load_dataset("OpenResearcher/OpenResearcher-Corpus-Gold-Doc", split="train")
row = ds[0]
print(row["query"])
print(row["answer"])
for doc in row["gold_docs"]:
print(doc["url"], doc["text"][:200])
```
## Related Resources
- Paper: [OpenResearcher: A Fully Open Pipeline for Long-Horizon Deep Research Trajectory Synthesis](https://arxiv.org/abs/2603.20278)
- Offline search corpus (gold docs + FineWeb distractors, embedded and indexed): [OpenResearcher/OpenResearcher-Corpus](https://huggingface.co/datasets/OpenResearcher/OpenResearcher-Corpus)
- Synthesized training trajectories: [OpenResearcher/OpenResearcher-Dataset](https://huggingface.co/datasets/OpenResearcher/OpenResearcher-Dataset)
- Code: [TIGER-AI-Lab/OpenResearcher](https://github.com/TIGER-AI-Lab/OpenResearcher)
## Citation
```bibtex
@article{li2026openresearcher,
title={{OpenResearcher: A Fully Open Pipeline for Long-Horizon Deep Research Trajectory Synthesis}},
author={Li, Zhuofeng and Jiang, Dongfu and Ma, Xueguang and Zhang, Haoxiang and Nie, Ping and Zhang, Yuyu and Zou, Kai and Xie, Jianwen and Zhang, Yu and Chen, Wenhu},
journal={arXiv preprint arXiv:2603.20278},
year={2026}
}
```
