SOURCES.md

# Input provenance and publication

The corpus is the YFCC-10M filtered-search dataset released for the NeurIPS 2023
Big-ANN competition. It contains precomputed vectors and tag metadata, not the
original photographs. The competition's [dataset table](https://big-ann-benchmarks.com/neurips23.html)
and [upstream README](https://github.com/harsha-simhadri/big-ann-benchmarks/blob/main/neurips23/README.md)
list this release under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
Credit belongs to the dataset creators and NeurIPS 2023 Big-ANN Filter Track
organizers; upstream references and original filenames are preserved below.

| Local file | Original source | Transformation |
| --- | --- | --- |
| `data/corpus/vectors.u8bin` | `https://dl.fbaipublicfiles.com/billion-scale-ann-benchmarks/yfcc100M/base.10M.u8bin` | Byte-identical copy under a task-local name |
| `data/corpus/metadata.spmat` | `https://dl.fbaipublicfiles.com/billion-scale-ann-benchmarks/yfcc100M/base.metadata.10M.spmat` | Byte-identical copy under a task-local name |
| `data/corpus/config.json` | Task-authored description of the upstream file headers | Records dimensions, dtypes, distance and tag counts |
| `data/validation/queries.jsonl` | Task-selected vectors and tag predicates derived from YFCC-10M inputs | Ten frozen public cases; query IDs are task-local |
| `data/validation/ground_truth.jsonl` | Task-author computation from the corpus and public queries | Exact filtered Top-10 with original-vector squared L2 |
| `data/example/*` | Task-authored small synthetic example | 64 documents and 17 functional queries, independent of the scored corpus |

The original query sources are `query.public.100K.u8bin` and
`query.metadata.public.100K.spmat` under the same upstream URL prefix. Query
selection originated in the original task's authoring workspace. Version 0.8.0
retains the original five cases per split and adds five from that workspace's
existing development/holdout candidate pool, using tag diversity before new
latency measurements. All twenty source rows, vectors and predicates are distinct;
each split has four single-tag, four double-tag and two triple-tag queries. Query
vectors are unchanged upstream rows. Exact labels for all cases were independently
recomputed by exhaustive float64 squared L2 over every eligible original vector.
The author-only provenance record retains source rows and selection details.

The original author recorded these full-corpus SHA-256 checksums; the migrated
copies have been checked against them:

- Vectors: `589030afcbcc44a50ff798cec935f0980815a02eb7075b224d4f7c12febe96bf`
- Metadata: `2f9c9a533fde31cb062edfdf5c9410f088bb5d7bf4a721ab98c47e3d2954287d`

`assets.json` records actual byte sizes and SHA-256 for all nine runtime input
files. The development dataset is `Cooki-e/search-swe-development`, under
`development/bingyu/1-x-1/`; use only the immutable revision recorded in
`assets.json` after publication.

Upstream CC BY 4.0 terms and attribution apply to the original dataset and its
derived inputs. The synthetic example and task-authored metadata are published
with the task author's authorization; no additional standalone license for
those original files was supplied. Their downstream license should be settled
with the author before official redistribution. The development dataset card
therefore points to these per-file terms instead of claiming one license for
every file. This document does not change the repository's code license.

Verifier-only queries and labels remain in `tests/data/` and are excluded from
the public input dataset and the agent image/mounts. Files under `tests/data/`
will still be visible to readers if this task code is published on GitHub;
runtime isolation is not a promise of private repository storage. Reference
solutions, private credentials and job outputs are excluded from publication.