Harden construction validation

construction/README.md

@@ -1,133 +0,0 @@
-# EVVE construction notes
-
-EVVE is a video-to-video specific-event retrieval benchmark introduced at
-CVPR 2013. The published core protocol contains 620 query videos, 2,375
-database videos, 13 events, and 135,213 positive query/database judgments. Its
-optional large-scale track adds 100,000 distractor videos.
-
-The original project page releases annotations, evaluation software, and
-descriptors and identifies the media by YouTube ID. Its legal note says that
-the software is BSD-licensed, descriptors are free, and copyright details for
-original videos remain on their YouTube pages. No dataset-wide video license is
-specified. The page does not state an explicit prohibition on redistribution.
-The constructor records this provenance without claiming that redistribution
-is authorized by EVVE or that video copyright was transferred.
-
-`s2vs-2023-video-ids.txt` freezes the video IDs present in the public S2VS EVVE
-feature artifact. Of those 2,410 IDs, 2,110 videos were reproducibly obtainable
-from public sources: 2,042 directly from their original YouTube IDs, 67 from
-exact Wayback Machine captures, and one from an exact Internet Archive item.
-`surviving-public-media-video-ids.txt` freezes that packaged set. Intersecting it
-with the checksum-pinned annotation file produces the fixed evaluation protocol:
-
-- 466 query videos;
-- 1,644 database videos;
-- all 13 events;
-- 86,925 binary relevance judgments.
-
-Filtering the S2VS layer removes 300 unavailable IDs: 38 queries and 262 database
-videos (81 positive and 181 other candidates), eliminating 13,864 qrels. The
-final protocol retains all 13 events, 958 positive database videos, 686 other
-candidates, zero query/corpus ID overlap, and 86,925 positive qrels. Every one
-of the 466 retained queries has at least one positive. `protocol-summary.json`
-records event-by-event original, S2VS-before-filter, and packaged-after-filter
-coverage.
-
-This is a reproducible surviving-public-media subset of EVVE, not the complete
-original benchmark. It has 154 fewer queries and 731 fewer database videos than
-the published core, and it omits the original 100,000 distractors. Published
-scores are therefore not directly comparable.
-
-Run a metadata-only integrity audit with:
-
-```bash
-uv run python scripts/data/evve/create_data.py --work-dir /tmp/evve-mteb
-```
-
-Download the exact frozen set reproducibly with yt-dlp. Downloads are resumable,
-limited to video at up to 360p, and never silently remove failures:
-
-```bash
-uv run python scripts/data/evve/create_data.py \
-  --work-dir /tmp/evve-mteb \
-  --download \
-  --workers 8
-```
-
-YouTube may require proof-of-origin tokens even for public, logged-out media.
-When that happens, install a guest PO-token provider alongside yt-dlp and pass
-its client/runtime settings without using account cookies. For example, with a
-compatible provider listening on its default local port:
-
-```bash
-uv run \
-  --with 'yt-dlp[default]==2026.7.4' \
-  --with 'bgutil-ytdlp-pot-provider==1.3.1' \
-  python scripts/data/evve/create_data.py \
-  --work-dir /tmp/evve-mteb \
-  --download \
-  --workers 4 \
-  --yt-dlp-arg=--js-runtimes \
-  --yt-dlp-arg=node \
-  --yt-dlp-arg=--extractor-args \
-  --yt-dlp-arg=youtube:player_client=mweb
-```
-
-`--yt-dlp-arg` is repeatable so current extractor requirements can be supplied
-explicitly and recorded with a construction run. Authentication cookies are
-optional and are never read unless `--cookies-from-browser` is provided.
-For a distributed or focused resume, `--download-id-file` accepts a newline-
-separated subset of the frozen IDs. It changes only the acquisition queue;
-packaging still validates the complete 2,110-video evaluation manifest.
-
-The constructor writes `download-failures.jsonl` and fails until all 2,110
-videos are present and contain decodable frames. Four YouTube IDs whose
-progressive format currently contains MP4 headers but no media payload are
-downloaded from separate video-only and audio streams and merged without
-transcoding. The complete frozen tree was also audited with the same TorchCodec
-10-frame uniform sampler that MTEB applies during evaluation; all 2,110 videos
-passed. On a resumed run it validates existing media
-only once, tries previously unattempted IDs before earlier transient and hard
-failures, and stops the pass immediately if YouTube starts returning its
-bot-confirmation gate. This preserves completed files and prevents temporary
-rate limits from being recorded as dataset attrition. `media-manifest.jsonl`
-then freezes each
-filename, original YouTube URL, actual acquisition URL, byte count, and SHA-256
-digest. `media-source-overrides.json` pins exact public archive captures and
-their checksums for source IDs that are no longer directly downloadable. To use
-an existing media tree, pass `--media-root`; files may be named
-`<youtube-id>.*` or placed beneath a directory named `<youtube-id>`.
-
-After validating the complete frozen set, push the standard MTEB `corpus`,
-`queries`, and `qrels` configurations plus the dataset card and manifests:
-
-```bash
-uv run python scripts/data/evve/create_data.py \
-  --work-dir /tmp/evve-mteb \
-  --repo-id Cerru02/EVVE \
-  --push
-```
-
-The Hub upload uses the currently authenticated Hugging Face identity and writes
-the immutable final revision to `hub_revision.txt`.
-
-Sources:
-
-- Original paper: https://openaccess.thecvf.com/content_cvpr_2013/html/Revaud_Event_Retrieval_in_2013_CVPR_paper.html
-- Archived project page: https://web.archive.org/web/20241102150758id_/http://pascal.inrialpes.fr/data2/evve/index.html
-- S2VS source revision: https://github.com/gkordo/s2vs/tree/03eb0c9a9a7455d132c4210b797fa6ef563b52ea
-- S2VS feature artifact: https://mever.iti.gr/s2vs/features/evve.hdf5
-- Surviving evaluator copy: https://github.com/fyang93/BURST/blob/c0f59d49ffe6b4069809794697554c8a1ce969d9/eval_evve.py
-
-The original evaluator uses trapezoidal AP and removes event-specific `null`
-videos from each ranking. It reports query-weighted `overall mAP` and the mean
-of 13 event-level mAP values as `avg-mAP`. Table 3 identifies the latter as
-`Event avg-mAP`. The MTEB task therefore retrieves all 1,644 database videos,
-preserves the per-query ignored-ID field (empty in the pinned annotation
-snapshot), and exposes event-balanced `evve_avg_map` as `main_score`;
-`evve_overall_map` is also reported. Ordinary pytrec
-`map_at_k` remains useful but is not identical to the source score.
-
-The original 100,000-video distractor collection is not publicly available and
-is not included. Together with core attrition, that makes scores on this frozen
-reduced corpus not directly comparable to source-paper results.

construction/create_data.py

@@ -517,8 +517,11 @@ def _video_is_decodable(path: Path) -> bool:
 
     try:
         from torchcodec.decoders import VideoDecoder  # type: ignore[attr-defined]
-    except ImportError:
-        return True
+    except ImportError as error:
+        raise RuntimeError(
+            "Media validation requires ffprobe or TorchCodec. Install FFmpeg or "
+            "MTEB's `video` extra before downloading or packaging EVVE."
+        ) from error
     try:
         duration = VideoDecoder(str(path)).metadata.duration_seconds
         return duration is not None and duration > 0
@@ -1151,7 +1154,6 @@ def push_dataset(
     for local_path, path_in_repo in (
         (card_path, "README.md"),
         (Path(__file__), "construction/create_data.py"),
-        (Path(__file__).with_name("README.md"), "construction/README.md"),
         (
             PROTOCOL_MANIFEST,
             f"construction/{PROTOCOL_MANIFEST.name}",