ybyby624/wiki-fixed

This repository contains the Wiki-Fixed corpus, presented in the paper Retrieval, Reward, and Training Protocols: What Matters in Training Search Agents?.

Code: https://github.com/YiboZhao624/SearchAgentReview

Description

The Wiki-Fixed corpus is based on the Wikipedia 2018 (Wiki-18) corpus and supplemented by the HotpotQA, 2WikiMultiHopQA, and Musique datasets. Compared to the original Wiki-18 corpus, this version contains 295,311 new documents which are critical for answering questions in these multi-hop reasoning datasets.

The corpus has been embedded using Qwen3-8B-Embedding. As the Qwen3-8B-Embedding is trained with Matryoshka Representation Learning (MRL), we provide the 4096-dimension version.

Sample Usage

If you are limited by computing resources, you can use the provided dimensionality reduction script to target a lower dimension (any dimension between 32 and 4096):

python convert_faiss_dim.py --input {your_faiss_manifest_path} --output {your_target_file_name_without_postfix} --target_dim {any_dim_between_32_and_4096} --normalize

After conversion, you can launch the search service with the converted index along with the original corpus. To enable the Qwen3-8B-Embedding's MRL capability, launch the vLLM server with the config --hf-overrides {"is_matryoshka": true}, and send requests with the dimension argument.

Citation

If you find this repo helpful, please cite:

@misc{zhao2026retrievalrewardtrainingprotocols,
      title={Retrieval, Reward, and Training Protocols: What Matters in Training Search Agents?}, 
      author={Yibo Zhao and Zichen Ding and Jiayi Wu and Zun Wang and Xiang Li},
      year={2026},
      eprint={2605.27881},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2605.27881}, 
}