Upload README.md with huggingface_hub

README.md

@@ -1,129 +1,21 @@
-# VMem-Bench: Memory-aware causal long video generation
-This Hugging Face dataset contains **textual gold / prompts only**. Source videos are not redistributed: obtain BlenderOpenMovies and LSMDC under their original terms. Annotations are CC BY 4.0 (see `LICENSE`).
-
 ---
-
-# Track A: Generator Gold-Replay
+license: cc-by-4.0
+pretty_name: VMem-Bench
 ---
 
-# Track B: Generator-in-the-loop
-
-Track B tests whether a causal long-video generator can carry visual memory all
-the way into generated pixels. The SUT only receives a chronological prompt
-stream; GT is scoring-only.
-
-## Frozen Status
-
-Frozen samples are listed in `freeze_manifest.json`.
-
-Current frozen sample:
-
-- `0001_lighthouse_keeper` (`trackB-0001-v1`, `gt_version=trackB-gt-2.0`)
-
-Do not edit frozen GT/prompt files in place. If a frozen sample must change,
-create a new freeze version and update hashes in `freeze_manifest.json`.
-
-## Files
-
-- `gt_source/<story>.json`: human-authored source layer.
-- `gt/<story>.json`: completed scoring GT generated by `complete_gt.py`.
-- `sut_prompts/<story>_<register>.json`: SUT-facing prompt stream generated by
-  `get_sut_prompts.py`.
-- `freeze_manifest.json`: immutable sample list and SHA256 hashes.
-
-## Current Scale
-
-The current validated authoring pool contains 50 long-horizon narrative cases.
-These are story-level cases, not single-prompt VBench-style cases: each case is
-a temporally ordered prompt stream with many segment-level memory probes.
-
-- Story-level cases: 50.
-- Segment-level cases: 6,161.
-- Segment count per story: mean 123.22, median 117, range 68-197.
-- Segment-count buckets: 50-80: 8 stories; 81-120: 19 stories; 121-160: 8
-  stories; 161-200: 15 stories.
-- Segment duration: 5 seconds.
-- Total scripted video duration: 30,805 seconds, or 8.56 hours.
-- Average story duration: 616.1 seconds, or 10.27 minutes.
-- Story duration range: 5.67-16.42 minutes.
-
-Hard-case annotations are segment-level. Across the 50-story pool, the current
-derived GT contains: lookalike disambiguation 723, state change 401,
-persist-state 1,878, count memory 279, false friend 250, deprecation avoidance
-578, reference-indirect 329, long-gap reappearance 476, and temporal reference
-265 tagged segments. Each story has at least 5 instances for every required
-hard-case family, and the current `sut_prompts/` files have zero prompt-rendering
-warnings.
-
-## Author Layer: `gt_source`
-
-Humans write only the semantic skeleton:
-
-- `entities`: stable entity IDs (`E1`, `E2`, ...) with `name`, `kind`, and either
-  `appearance` or state-specific `states`.
-- `state_machines`: ordered state labels for stateful entities.
-- `lookalike_pairs`: explicit pairs and distinguishing features.
-- `scenes`: chronological scenes with `present`, per-segment `actions`, optional
-  per-action `present`, optional `lookalike_present`, and semantic events:
-  `state_change` or `remove`.
-
-Do not hand-write memory labels such as `recall_after_gap`, `forbidden`, or
-probe buckets. They are derived by `complete_gt.py`.
-
-## Completed GT: `gt`
-
-`complete_gt.py` derives:
-
-- per-entity `op`: `introduce`, `recall`, `recall_after_gap`, `transform`,
-  `persist`;
-- gap and last-seen metadata;
-- `memory_probes` capability tags;
-- `forbidden` entries for removals and absent lookalikes;
-- state payloads for transform/persist cases;
-- summary counts and balance diagnostics.
-
-The scorer reads the v2 schema: `cast`, `forbidden`, `memory_probes`, and
-`lookalike_active`. It derives present sets and decoys at score time.
-
-## SUT Prompts
-
-`get_sut_prompts.py` creates the only SUT-facing input. For `name_anchored`:
-
-- `introduce`: inject the entity's initial appearance after the name;
-- `transform`: inject the new state's appearance after the name;
-- `recall`, `recall_after_gap`, `persist`: keep only the bare name;
-- `forbidden`: never inject appearance, even if the name appears indirectly.
-
-The prompt stream must contain no entity IDs, probe labels, forbidden labels, or
-scoring fields.
-
-## Freeze Checklist
-
-A sample can be frozen only if all checks pass:
-
-1. `gt_version` is `trackB-gt-2.0`.
-2. `gt/<story>.json` regenerates byte-identically from `gt_source/<story>.json`.
-3. `sut_prompts/<story>_name_anchored.json` regenerates byte-identically from GT.
-4. SUT prompt warnings are zero.
-5. SUT prompts contain no labels (`E#`, `memory_probes`, `forbidden`,
-   `present_required`, `decoy`, etc.).
-6. `complete_gt.py` summary has no warnings and enough capability coverage for
-   the target hard cases.
-7. SHA256 hashes for source, GT, prompts, and generation scripts are recorded in
-   `freeze_manifest.json`.
-
-## Rebuild Commands
+# VMem-Bench (dataset)
 
-```bash
-python3 benchmarks/VMem-Bench/assets/trackB/scripts/complete_gt.py --story 0001_lighthouse_keeper
-python3 benchmarks/VMem-Bench/assets/trackB/scripts/get_sut_prompts.py --story 0001_lighthouse_keeper
-```
+Textual gold and SUT prompts for [VMem-Bench](https://github.com/Suchenl/VMem-Bench). **No source videos.**
 
-Use temporary output directories for audits, then compare with `cmp -s` before
-updating any frozen artifact.
+| Split | Contents |
+|---|---|
+| `trackA/` | Per-movie `gold/{entity_registry,chunk_annotations,chunk_index}.json` (Blender Open Movies + LSMDC titles) |
+| `trackB/` | 50 long-horizon stories: `gt`, `gt_source`, `sut_prompts` (zh/en) + `freeze_manifest.json` |
 
-## SHA256 In One Sentence
+- Code / scoring: https://github.com/Suchenl/VMem-Bench
+- Method: https://github.com/Suchenl/MemStrata
+- License of **these JSON files**: CC BY 4.0 (`LICENSE`)
+- LSMDC **videos**: apply at the [LSMDC download page](https://sites.google.com/site/describingmovies/download), stitch to `LSMDC_Videos_Stitched/<movie_id>.mp4`; cite Rohrbach et al., IJCV 2017. Full steps: [VMem-Bench `docs/DATA.md`](https://github.com/Suchenl/VMem-Bench/blob/main/docs/DATA.md)
+- Blender **videos**: official project pages / `scripts/prepare_blender.sh` for BBB; same `DATA.md`
 
-SHA256 is a deterministic file fingerprint. The same bytes always produce the
-same 64-character hash; any content change produces a different hash, so it is
-useful for proving a frozen GT or prompt file did not silently change.
+See `trackA/README.md` and `trackB/README.md` in this dataset for file schemas.