video video |
|---|
Video Failure Benchmark
Video Failure Benchmark is a pilot research dataset for fine-grained diagnosis of generated-video failures: what went wrong, which object or region is involved, and when the evidence occurs. It organizes observations into 33 failure subclasses across eight families, linking descriptive annotations with a systematic video–subclass screening matrix.
The benchmark supports research on failure discovery, subclass classification, and temporal diagnosis, including failures involving geometry, motion, physical interactions, and causal consistency. These directions are relevant to physical-world understanding and evaluation, and this anonymous release accompanies the poster submission to the PhysWorldAI workshop at NeurIPS 2026.
The current release includes 96 generated videos, 278 accepted annotation records, and 3,168 video–subclass matrix entries.
Dataset: anon482915/video-failure-benchmark
Dataset at a glance
| Item | Current release |
|---|---|
| Generated videos / JSONL sample records | 96 / 96 |
Text-to-video (T2V) samples |
87 |
Image-to-video (I2V) samples |
9 |
| Failure families / subclasses | 8 / 33 |
| Accepted annotation records | 278 |
| Video–subclass matrix entries | 3,168 = 96 × 33 |
| Video–subclass cells with accepted annotations | 240 |
| Total video duration | Approximately 13.1 minutes |
Each video contains between one and eight accepted annotation records. Multiple records may occupy the same video–subclass cell; annotation records, positive cells, and distinct events are different counting units.
The MP4 files vary in resolution and frame rate, with durations ranging from approximately 0.8 to 22.31 seconds. The release is approximately 467 MB.
Repository structure
.
├── README.md
├── .gitattributes
├── videos/
│ ├── VID-001.mp4
│ ├── ...
│ └── VID-096.mp4
├── annotations/
│ └── labels.jsonl
├── metadata/
│ ├── review_matrix.csv
│ ├── human_error_review_template.csv
│ ├── human_example_review_template.csv
│ ├── timing_completion.csv
│ ├── annotation_corrections.json
│ ├── annotation_rechecks.json
│ ├── human_adjudications.json
│ ├── review_candidates.json
│ └── gt_release_audit.json
├── REVIEW_PROTOCOL.md
├── HUMAN_REVIEW_GUIDE.md
└── REVIEW_GUIDE.md
| File | Purpose |
|---|---|
| annotations/labels.jsonl | Current accepted annotations: one JSON object per video, with nested error records. |
| metadata/review_matrix.csv | All 96 × 33 video-subclass pairs, including applicability and screening state. |
| metadata/human_error_review_template.csv | 278 prefilled annotation rows with blank fields for independent reviewer decisions and timing corrections. |
| metadata/human_example_review_template.csv | Three example-review template slots per subclass (99 total). |
| metadata/timing_completion.csv | Historical timing notes for 31 records, including observation windows and full-clip assessment scopes. |
| metadata/annotation_corrections.json | Correction history, previous descriptions, withdrawn claims, and mappings from redundant IDs to retained IDs. |
| metadata/annotation_rechecks.json | Annotation recheck and resolution history. |
| metadata/human_adjudications.json | Returned human decisions and later application notes; historical decisions remain preserved. |
| metadata/review_candidates.json | Candidate annotation awaiting adjudication. |
| metadata/gt_release_audit.json | Release counts, cleanup scope, and explicit limits on what the audit certifies. |
| REVIEW_PROTOCOL.md | Current screening conventions, temporal interpretation, and correction history. |
| HUMAN_REVIEW_GUIDE.md / REVIEW_GUIDE.md | Operational subclass definitions and independent review instructions. |
Use labels.jsonl for current accepted positives and review_matrix.csv for coverage state. Correction logs and timing histories can retain removed or superseded IDs; do not concatenate them into the accepted labels. IDs are stable and have not been renumbered after removals.
Quick start
Download the release with the Hugging Face Hub client:
python -m pip install huggingface_hub
hf download anon482915/video-failure-benchmark \
--repo-type dataset \
--local-dir video-failure-benchmark
For reproducibility, you can pin the inspected data snapshot with --revision def191b0b94b1bd7059b8e67381f83c1dd4e757a (a commit predating this README). See the Hub download documentation for details.
The following example uses only the Python standard library and runs from the directory containing the downloaded folder:
import csv
import json
from pathlib import Path
root = Path("video-failure-benchmark")
with (root / "annotations/labels.jsonl").open(encoding="utf-8") as f:
samples = [json.loads(line) for line in f if line.strip()]
with (root / "metadata/review_matrix.csv").open(
newline="", encoding="utf-8"
) as f:
matrix = list(csv.DictReader(f))
check_ids = {
(row["category"], row["subcategory"]): row["check_id"]
for row in matrix
}
assert len(samples) == 96
assert len(matrix) == 96 * 33
assert sum(len(sample["errors"]) for sample in samples) == 278
assert all((root / sample["video_path"]).is_file() for sample in samples)
sample = next(s for s in samples if s["sample_id"] == "VID-004")
print(sample["prompt"])
print(root / sample["video_path"])
for error in sample["errors"]:
check_id = check_ids[(error["category"], error["subcategory"])]
print(
error["error_id"], check_id,
(error["start_sec"], error["end_sec"]),
error["abnormal_fact"],
)
For this sample, the accepted records describe a misaligned contact shadow (P04, 4.7–6.6 seconds) and an unstable grid texture (A02, 3.2–6.4 seconds). Video paths are relative to the dataset root, not to the annotations/ directory.
Annotation schema
annotations/labels.jsonl contains 96 lines, not 278 lines. The 278 error records are nested in the per-video errors lists.
Per-video fields
| Field | Type | Meaning |
|---|---|---|
sample_id |
string | Stable video identifier, such as VID-004. |
video_path |
string | Relative path to the released MP4. |
source_video_url |
null in this release | Source-link field; all 96 values are null in the anonymous export. |
task |
string | T2V or I2V. |
prompt |
string | Released English prompt or scene-description text. |
annotation_origin |
string | curated_annotation for all current samples. |
review_status |
string | Current sample-level curation status (confirmed in this release). |
errors |
list of objects | Current accepted error records for this video. |
Per-error fields
| Field | Type | Meaning |
|---|---|---|
error_id |
string | Stable annotation ID, such as VID-004-E01. Gaps are expected after corrections. |
category |
string | One of the eight family names below. |
subcategory |
string | Machine-readable subclass name. |
location |
string | Human-readable object, region, timing, and evidence-scope description. |
abnormal_fact |
string | Concrete observed failure claim. |
correct_state |
string | Constraint or expected behavior against which the observation is judged. |
expected_result |
string | Description of the intended visible outcome, not a repaired target video. |
start_sec, end_sec |
number | Bounds in seconds on the original video timeline; read together with location. |
spatial_annotation |
null in this release | Reserved field for spatial annotations. |
The short taxonomy code (A01, G03, etc.) is stored as check_id in the matrix and human error review template. It is not a field in the nested JSON errors; join using (category, subcategory) as in the example above.
Failure taxonomy
Counts are accepted annotation records, not independent events, positive video counts, or failure rates. The full operational definitions appear in HUMAN_REVIEW_GUIDE.md.
Family (category) |
Subclasses | Records |
|---|---|---|
appearance |
3 | 16 |
temporal_consistency |
4 | 39 |
motion |
4 | 23 |
geometry |
8 | 52 |
physics |
4 | 42 |
semantic_instruction |
4 | 70 |
causal_logic |
3 | 23 |
input_preservation |
3 | 13 |
| Total | 33 | 278 |
| Check ID | Subclass (subcategory) |
Records |
|---|---|---|
| A01 | blur_detail_loss |
3 |
| A02 | texture_artifact |
5 |
| A03 | ghost_boundary_artifact |
8 |
| T01 | appearance_disappearance |
20 |
| T02 | identity_drift |
4 |
| T03 | attribute_shape_drift |
12 |
| T04 | flicker |
3 |
| M01 | action_freeze |
6 |
| M02 | jitter |
3 |
| M03 | trajectory_speed_discontinuity |
7 |
| M04 | action_coordination |
7 |
| G01 | anatomical_structure |
9 |
| G02 | shape_deformation |
9 |
| G03 | topology_fusion_split |
14 |
| G04 | pose_structure |
7 |
| G05 | spatial_projection |
3 |
| G06 | cross_view_surface_structure |
3 |
| G07 | relative_size_scale_anomaly |
4 |
| G08 | scene_layout_inconsistency |
3 |
| P01 | contact_support_collision |
11 |
| P02 | force_dynamics |
13 |
| P03 | material_fluid_conservation |
8 |
| P04 | optics |
10 |
| S01 | entity_attribute_count |
13 |
| S02 | action_goal |
49 |
| S03 | relation_binding |
5 |
| S04 | specified_order |
3 |
| C01 | cause_effect |
5 |
| C02 | event_dependency |
4 |
| C03 | state_relation |
14 |
| R01 | reference_identity |
3 |
| R02 | reference_appearance_structure |
5 |
| R03 | reference_layout_unedited_region |
5 |
Subclass boundaries require the observed fact and scene context. For example, T02 changes recognizable identity, while T03 changes an attribute while identity persists. Prompt-permitted transformations, ordinary occlusion, perspective changes, and expected motion blur should not automatically be labeled as failures.
Screening and human review
metadata/review_matrix.csv records a screening state for every video–subclass pair, along with the sample, task, category, subclass, and review status.
Matrix status |
Cells | Interpretation |
|---|---|---|
confirmed_error |
240 | At least one accepted annotation exists for this video–subclass pair. |
not_confirmed |
2,587 | Screening recorded no accepted positive for this pair; this is not a verified negative. |
not_applicable |
341 | The subclass check is outside the recorded scope for this sample. |
The separate review_status column records confirmed (240 cells), checked (2,695), and checked_with_limits (233). These values describe review state rather than a binary ground-truth label.
The screening process used visual inspection of frame sequences sampled across each clip. The accompanying draft discloses language-model assistance in taxonomy organization, visual screening, candidate descriptions, and record normalization. The supplied templates support independent review and adjudication, in which reviewers can record present, absent, uncertain, or not_applicable and propose timing corrections. See REVIEW_PROTOCOL.md and HUMAN_REVIEW_GUIDE.md for the full workflow.
Suggested evaluation workflow
The accompanying manuscript proposes three complementary tasks: failure discovery, subclass classification, and temporal localization. Researchers can use the released videos, annotation labels, screening matrix, and review templates to study these tasks.
- Define the evaluation subset. Pin a dataset revision and choose the relevant samples, failure classes, and supported input modalities. If creating train/validation/test partitions, group shared prompts and near-duplicate content before splitting.
- Run diagnosis on the videos. Provide the evaluator with the video, prompt, and taxonomy. Keep accepted error descriptions and correction logs out of the input when testing failure discovery.
- Match and classify detected failures. Compare detected failures against reviewed annotations by the affected object and observed abnormality, then assess the predicted subclass for matched events. Review overlapping descriptions before counting independent events.
- Evaluate temporal diagnosis. Compare reported intervals with annotation timing where appropriate, validating the reference bounds and distinguishing visible-event intervals from broader assessment windows.
- Report evaluation conditions. Document sampling procedures, excluded cases, matching criteria, and any additional human adjudication. Review unmatched predictions before treating them as false positives.
Scope and release notes
This release is designed as a curated diagnostic research resource, rather than a representative estimate of how often particular video-generation failures occur. The following details are useful when designing evaluations:
- Screening coverage:
not_confirmedmatrix entries are not verified negatives, and the screening process does not establish exhaustive frame-by-frame coverage or independent inter-annotator consensus. - Temporal and spatial annotations: All accepted records have numerical time bounds, although some are approximate observation or full-clip assessment windows rather than frame-exact event boundaries. In
VID-006, a few stored end times slightly exceed the MP4 container duration; validate such cases before temporal scoring. The release does not provide populated spatial masks, boxes, or points. - Input modalities: The nine
I2Vsamples retain reference-dependent annotations, but the original reference images and camera-control inputs are not packaged here. SomeVID-073annotations concern audible events, so they require audio for review. - Additional materials: Official benchmark splits, a scoring implementation, and the separate problem-centric document referenced in the review guides are not part of this snapshot. A Hugging Face viewer-generated
trainsplit is not an official evaluation split.
The current statistics reflect the 2026-10-29 annotation cleanup, which revised the draft's earlier 297 records to 278 accepted records: 15 redundant records were removed, three unsupported records were withdrawn, and one record was retained as an unadjudicated candidate (VID-051-E01). Historical decisions and superseded IDs remain in the metadata for traceability.
Provenance, licensing, and citation
The accompanying draft describes the video collection as curated from public benchmarks and official generated-video examples. The anonymous export uses neutral sample IDs and leaves source_video_url null. A per-asset attribution/license manifest and an explicit dataset license grant are not provided in this snapshot; users should verify applicable source rights before redistributing or reusing the videos.
When using this release, cite Video Failure Benchmark and its anonymous Hugging Face repository, and report the dataset revision and evaluation subset. Author information and a permanent paper identifier are intentionally omitted from this anonymous poster-submission release.
- Downloads last month
- 302