Dataset Viewer
Duplicate
The dataset viewer is not available for this split.
The info cannot be fetched for the config 'default' of the dataset.
Error code:   InfoError
Exception:    KeyError
Message:      'default'
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/split/first_rows.py", line 224, in compute_first_rows_from_streaming_response
                  info = get_dataset_config_info(path=dataset, config_name=config, token=hf_token)
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 268, in get_dataset_config_info
                  builder = load_dataset_builder(
                      path,
                  ...<6 lines>...
                      **config_kwargs,
                  )
                File "/src/services/worker/src/worker/utils.py", line 464, in safe_load_dataset_builder
                  config_data_files = builder_cls.builder_configs[name].data_files
                                      ~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^
              KeyError: 'default'

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

Repo-Premium

Repository-level code retrieval benchmark accompanying Embeddings vs. Grep: Rethinking the Dense Advantage in Code Search.

Given a natural-language issue or pull-request description, rank files in the matching repository snapshot. This release contains queries, file relevance labels, fixed evaluation membership, and provenance. Repository source code and model weights are obtained separately.

Paper · Code and evaluation scripts · Paper artifact archive

Load

from datasets import load_dataset

core = load_dataset("hq-bench/repo-premium", "core", split="test")
raw_2023 = load_dataset("hq-bench/repo-premium", "raw_pre_2023", split="test")
lockbox = load_dataset("hq-bench/repo-premium", "lockbox", split="test")

example = core[0]
print(example["query"])
print(example["repo"], example["corpus_commit"])
print(example["gold_files"])

These are evaluation pools, not random train/test partitions. Preserve the original pool boundaries. The lockbox overlaps extension pools; do not concatenate it with them as additional independent examples.

Evaluation configurations

Configuration Records Description
core (default) 1,265 Four final Python pools below
cur_pre 472 Curated SWE-bench issue statements
raw_post 288 Raw PR descriptions from 2026
raw_pre_2023 138 Raw 2023 PRs, evaluated on 2024-Q1 corpora
raw_pre_2022 367 Raw 2022 PR descriptions
extensions 1,560 Five extension pools below
cross_language 458 PR queries across JS/TS, Go, and Rust; 443 have nonempty source-only labels
multi_swe_bench 134 Curated JS/TS and Rust issues
heldout_python 139 Held-out Python queries with nonempty source labels
matched_replication 720 Raw and curated views of 360 tasks; 351 tasks / 702 views have scored source labels
deepswe 109 Task specifications over 88 repositories
lockbox 539 Fixed preregistered membership across 91 repositories; original release labels

The original 7,431 source records are also available in source_pools, split source. They include mining candidates, rejected records, and query variants. They are not an evaluation set or additional independent examples. source_record_json preserves each original published record, including its original filtering fields. The 500-record curated source pool is distinct from the final 472-query evaluation pool.

Fields and labels

Field Meaning
id Unique row identity including pool, repository, and original query ID. Source-pool rows also include the original file and line.
query_id Original query ID; matched views retain ::raw / ::curated suffixes.
pool, repo, language Query pool, upstream repository, and language or language group.
query Retrieval query from the scrubbed archive, with the pool's evaluated boilerplate preprocessing applied. Matched-replication views and DeepSWE task specifications retain the historical verbatim input. No generated anchors are added. Source-pool queries retain their published text.
gold_files Relevant paths under the row's label_policy.
gold_files_source Indexed non-test source-file label alternative.
gold_files_as_released Original labels before the source-only view; raw core rows preserve original pre-harmonization touched-file labels where available.
label_policy indexed_non_test_source for evaluation pools except the original-label lockbox; as_released for lockbox and source records.
eligible_source_labels Whether the source-file label set is nonempty. This alone does not certify a mining candidate passes evaluation filters.
scored_in_paper Whether the record belongs to the cached evaluation under its primary label policy. Source-pool records are marked false.
base_commit Task/PR base commit; not automatically the indexed shared corpus commit.
corpus_id, corpus_commit, corpus_status Indexed corpus identity, exact commit when recovered, and verification status.
source_path, source_record_json Pinned archive provenance and the underlying record as JSON.

For source-only analyses, exclude rows with eligible_source_labels == False; retain the original membership and label policy when reproducing the preregistered lockbox result. The source-only lockbox has 525 eligible rows, while the original-label lockbox has 539. The matched replication includes all 360 selected tasks; use scored_in_paper to select the 351 scored tasks. Do not score empty source gold sets as retrieval failures.

Corpus reconstruction

metadata/corpora.json specifies the 20 shared core-Python corpora (ten original snapshots and ten 2024-Q1 snapshots), upstream URLs, exact commits, and indexing rules. metadata/corpus_files.parquet records every indexed .py path, content SHA-256, Git blob ID, and file mode. The export checks the evaluated local file contents against Git objects, including symlink target contents. Tests remain in the indexed corpus even when excluded from relevance labels.

Clone the row's upstream repository and check out corpus_commit, then index the paths listed for its corpus_id. metadata/checksums.json verifies the release data and metadata files. metadata/provenance.json pins the source files to the archived release.

Legacy extension limitation: exact shared snapshot SHAs for some extension repositories were not recorded in the old release. Those rows have an empty corpus_commit and corpus_status="not_recorded_in_paper_archive". Do not substitute the PR's base commit and claim the same evaluation corpus. Matched-replication and DeepSWE rows record their task base commits, but their corpora were not reconstructed by this export. metadata/lockbox_protocol.json and metadata/lockbox_driver.tsv retain the exact portable lockbox runner settings, including corrected extension lists. Cached results remain in the paper archive. This cleanup preserves those limitations explicitly rather than inventing missing provenance.

Paper reproduction and migration

The full experiment tree is preserved at the immutable paper-artifacts tag, pointing to commit 97c0c94062634d8266a7238546169123a8ef2125. It contains the original release/, retrieval caches, analyses, sweeps, trajectories, and auxiliary inputs. Legacy paths are available at that revision, not on the new benchmark-focused main.

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="hq-bench/repo-premium",
    repo_type="dataset",
    revision="97c0c94062634d8266a7238546169123a8ef2125",
    local_dir="paper-data",
)

Set REPO_PREMIUM_DATA to the absolute path of paper-data when running the paper's table-reproduction commands. The new Parquet layout is for dataset users; the archived layout remains compatible with the experiment scripts.

The final raw_pre_2023 pool comes from the 138-query qrels_h24/cellRA inputs joined against out24/dense_cellRA_c1000. The earlier 135-query copy is retained in the archive and is not used for the new final pool. No new query sampling or retrieval runs were used to construct this release.

Privacy and licenses

Queries and labels come from the previously published, scrubbed archive. Personal emails were redacted during that release; technical email-shaped tokens and @-mentions may remain. Reported retrieval scores were computed before release-time redaction, so rerunning on scrubbed text can change rankings.

The dataset compilation is licensed under CC-BY-4.0. Source-derived issue text and code remain subject to their upstream licenses; the compilation license does not replace them. Repository URLs in the rows and corpus metadata provide source attribution. Obtain and retain the license at the pinned upstream commit when reconstructing a corpus. The benchmark code is Apache-2.0.

Citation

@article{xue2026embeddingsgrep,
  title={Embeddings vs. Grep: Rethinking the Dense Advantage in Code Search},
  author={Xue, Siqiao and Zhang, Ziyin and Liao, Zihan and Qin, Jin and Mu, Yixiang and Yu, Hang},
  journal={arXiv preprint arXiv:2611.20455},
  year={2026},
  doi={10.48550/arXiv.2611.20455},
  url={https://arxiv.org/abs/2611.20455}
}
Downloads last month
116

Collection including hq-bench/repo-premium

Paper for hq-bench/repo-premium