task: add InsAVE-80K composed video retrieval

[myang333]

Adds InsAVE80KVT2VRetrieval (vt2v, beta): query = source video + editing instruction → target video.

Closes #5146

New MOEB construction, not an official benchmark

Derived from the official InstructAV2AV 1,000-pair generation-evaluation split. The paper (arXiv:2605.18467) scores generation quality there (FVD, FAD, PEAVS, Sync-C/D) and defines no retrieval evaluation and no reference scores — nothing here reproduces a published number.

Construction

• Queries: the 1,000 official pairs — source clip + forward instruction, verbatim (incl. <S>/<E> markers).
• Corpus: all 2,000 released eval clips. Nothing added or dropped.
• Source clips remain hard negatives, deliberately not excluded, per CIRRIT2IRetrieval/FashionIQIT2IRetrieval. The instruction is what separates the edited clip from the unedited one.
• Qrels: the declared target plus any clip byte-identical (SHA-256) to it. The split has 60 add/remove reverse couples where one row's source file is the next row's target file, so the gold appears under two ids: 880 queries have 1 gold, 120 have 2 (1,120 total). Similarity never qualifies, only identical bytes.

Data

Pinned revision 8ba8ccaad1d8f97b08e218dae1f2b439bd3d2289. Only eval.csv and the single 1.55 GB eval shard are downloaded; the 139 GB training data is never touched.

KNOWN_ISSUES

The duplicate_video entry records an intrinsic property of the release: the 2,000 files hold 1,872 unique video contents. Not masked — the qrel rule means no model is penalised for returning identical bytes under the other id.

License

not specified: the repo declares MIT, but clips derive from YouTube/MovieBench/Condensed Movies/VGGSound and the card disclaims underlying media rights, as for YouCook2/ActivityNetCaptions.

Validation

mteb/baseline-random-encoder: ndcg@10 0.0029, R@1 0.0, R@1000 0.473 (≈ chance). Offline loader tests cover collision and schema guards.

[Samoed]

You shoudn't add tests for tasks loading

[Samoed]

This seems complicated. Can you reupload this task with fixed format? You can do this with task.push_dataset_to_hub

[Samoed]

Suggested change

	"""Composed video+text retrieval derived from the InsAVE-80K evaluation split.
	This is a **new MOEB retrieval construction**, not an official InstructAV2AV
	benchmark. InstructAV2AV uses these 1,000 pairs to score *generation* quality
	(FVD, FAD, PEAVS, Sync-C/D and similar); neither the paper nor the official code
	defines or reports any retrieval evaluation. The retrieval framing, the candidate
	pool and the relevance judgements below are introduced here and should not be
	compared against numbers reported in the paper.
	Construction, derived entirely from the pinned release:
	* Queries are the 1,000 official pairs: the source clip plus the forward
	`instruction`, kept verbatim (including the `<S>`/`<E>` speech markers and the
	13 rows whose instruction text contains upstream pipeline artefacts).
	* The corpus is all 2,000 released evaluation clips, i.e. the complete media
	contents of the eval shard. Nothing is added and nothing is dropped.
	* A query's own source clip therefore stays in the pool as a hard negative and is
	deliberately **not** excluded, following the `CIRRIT2IRetrieval` and
	`FashionIQIT2IRetrieval` precedent. The instruction is what separates the edited
	clip from the unedited one.
	* Relevance is expanded over **exact byte identity only**. The split contains 60
	add/remove reverse couples, where one row's source *file* is byte-for-byte the
	next row's target *file*, so the gold video is present in the corpus under two
	ids. Both ids are marked relevant so a model is not penalised for returning the
	identical bytes under the other name. This affects 120 of the 1,000 queries,
	which carry 2 gold ids; the remaining 880 carry 1. Visual or semantic similarity
	never qualifies -- only an identical SHA-256 digest does -- so a source clip can
	never become relevant merely by resembling the target.
	The closest upstream precedent for content-identity handling is `GreekCivicsQA`
	and `XQuADRetrieval`, which hash document content to derive ids and thereby
	collapse duplicates into a single document. That approach is not used here because
	the corpus is meant to preserve every released file as its own candidate; qrel
	expansion encodes the same fact without dropping ids.
	"""

[github-actions]

@github-copilot review

Please review this PR against the guidelines in .github/instructions/dataset-pr.instructions.md. Only report actionable issues. Include file/line references and a concrete suggested fix where possible.