Fix some mistaken answers and remove redundant questions

README.md

@@ -37,7 +37,7 @@ dataset_info:
   - config_name: natural_video
     splits:
       - name: train
-        num_examples: 1813
+        num_examples: 1726
   - config_name: easyhuman
     splits:
       - name: train
@@ -53,7 +53,7 @@ Each row is a `(stream, probe, label)` tuple: the model sees a long input stream
 | --- | --- | --- |
 | `text` | 7,640 | synthetic substream questions, text modality, L=8…4096 |
 | `synthetic_video` | 6,065 | the same streams rendered as video |
-| `natural_video` | 1,813 | 1,028 EPIC-Kitchens-100 + SoccerNet clips, L=8…1024, with two question sets over them — see `question_set` below |
+| `natural_video` | 1,726 | 1,015 EPIC-Kitchens-100 + SoccerNet clips, L=8…1024, with two question sets over them — see `question_set` below |
 | `easyhuman` | 672 | 3-belt EasyHuman: 224 video rows (L=256) and 448 text rows (224 at L=256, 224 at L=1024). `modality` distinguishes them |
 
 ## Loading
@@ -86,13 +86,13 @@ means the row has no media. Filter with `media_included != False`, not
 The key is `(question_id, length_L, entropy_band, question_variant)` — plus
 `modality` in `easyhuman`, where each question appears once as text and once as video.
 
-`natural_video` holds two question sets over the same 1,028 clips, told apart by
+`natural_video` holds two question sets over the same 1,015 clips, told apart by
 `question_set`:
 
 | `question_set` | rows | what it asks |
 |---|---|---|
-| `designed` | 1,028 | the one question the clip was built for, at every length L=8…1024 |
-| `density_topup` | 785 | extra mined questions about *other* classes, added for the density analysis, on the 403 codec-banded clips (L=64…1024) |
+| `designed` | 1,007 | the one question the clip was built for, at every length L=8…1024 |
+| `density_topup` | 719 | extra mined questions about *other* classes, added for the density analysis, on 377 codec-banded clips (L=64…1024) |
 
 They are disjoint: no shared `question_id`, and no clip is asked the same question
 twice. Filter on `question_set` to recover either subset.
@@ -101,12 +101,15 @@ Natural-video-specific columns:
 
 | column | meaning |
 |---|---|
-| `density_band` | `low` / `high`, codec distortion at a fixed 32 kbit/frame budget, one global cutoff. `null` on the 625 `designed` rows at L=8/16 and on unbanded clips — those buckets predate the codec protocol |
+| `density_band` | `low` / `high`, codec distortion at a fixed 32 kbit/frame budget, one global cutoff. `null` on the 613 `designed` rows at L=8/16 and on unbanded clips — those buckets predate the codec protocol |
 | `event_start_frame`, `event_end_frame` | **integer frame indices** of the clip's designed event — index the decoded clip directly |
-| `event_clip_start`, `event_clip_end` | the same instants in seconds. Clips are 1 fps with frame count = duration, so the frame is `floor(seconds)`. Populated on 882 of 1,813 rows; `null` where the clip has no localised event |
+| `event_clip_start`, `event_clip_end` | the same instants in seconds. Clips are 1 fps with frame count = duration, so the frame is `floor(seconds)`. Populated on 871 of 1,726 rows; `null` where the clip has no localised event |
 | `event_timing` | `measured` (EPIC: extent from narration) or `centered_by_construction` (SoccerNet: labels give an instant, so the clip is centred on it with a nominal 2 s window) |
 | `clip_sha256` | sha256 of the exact mp4 published at `media_config`/`video_path`. `null` only where the media is NDA-gated |
-| `label_source` | how the label was derived: `designed` for the designed questions, one of `narration_margin` / `whole_video_absence` / `labels_v2_window` / `labels_v2_window_absence` for the mined ones |
+| `label_source` | how the label was derived: `designed` for the designed questions, one of `narration_margin` / `whole_video_absence` / `labels_v2_window` / `labels_v2_window_absence` for the mined ones; `annotation_corrected` where the answer was re-checked against the source annotations and corrected |
+| `in_density_comparison` | Whether the paper's natural-video density comparison uses this question (the accuracy-versus-compute figure and the band-composition and band-contrast tables). At L=512 and L=1024 the question subsets differ by model; true means at least one model. |
+| `in_calibration` | Whether the paper's natural-video calibration experiment uses this question (the coverage and selectivity figures and the calibration table). |
+| `in_rag_comparison` | Whether the paper's retrieval (RAG) versus full-video comparison uses this question. |
 | `query_class` | the class this question asks about |
 | `question_time` | when the question is put, in clip seconds. Always `length_L`: the whole stream is shown first |
 | `source_provenance` | JSON, on every row: `source_video_id`, `clip_start/end_in_source_seconds`, `event_position/end_in_source_seconds`, `event_type`, `slowdown_factor`, `anomaly_label`, plus SoccerNet match/half/path/entry-id |

metadata.json

@@ -74,7 +74,7 @@
   "name": "Anonymous Authors"
  },
  "datePublished": "2026-05-03",
- "rai:dataCollection": "Substream Recollection has two components. (1) Synthetic streams over a fixed alphabet of 16 symbols, generated by a variable-order n-gram process (rule lengths 1-8) and resampled until the realized stream falls within a target empirical Lempel-Ziv entropy band (low ~0.7 bits/symbol, medium ~1.9 bits/symbol, max-entropy = 4.0 bits/symbol = uniform i.i.d.). The 'EasyHuman' diagnostic split is generated separately by hand-specified pattern templates (repeating motifs and counting sequences) with controlled deviations. (2) A natural-video subset of 1,028 clips drawn from EPIC-Kitchens-100 (egocentric cooking videos) and SoccerNet (sports broadcasts) at L in {8, 16, 64, 128, 512, 1024}; EPIC-Kitchens-100 derived clips ship in this repo and SoccerNet rows ship as provenance metadata only.",
+ "rai:dataCollection": "Substream Recollection has two components. (1) Synthetic streams over a fixed alphabet of 16 symbols, generated by a variable-order n-gram process (rule lengths 1-8) and resampled until the realized stream falls within a target empirical Lempel-Ziv entropy band (low ~0.7 bits/symbol, medium ~1.9 bits/symbol, max-entropy = 4.0 bits/symbol = uniform i.i.d.). The 'EasyHuman' diagnostic split is generated separately by hand-specified pattern templates (repeating motifs and counting sequences) with controlled deviations. (2) A natural-video subset of 1,015 clips drawn from EPIC-Kitchens-100 (egocentric cooking videos) and SoccerNet (sports broadcasts) at L in {8, 16, 64, 128, 512, 1024}; EPIC-Kitchens-100 derived clips ship in this repo and SoccerNet rows ship as provenance metadata only.",
  "rai:dataCollectionType": "Mixed: programmatically generated synthetic streams (text and rendered video) plus curated derivatives of two existing public video datasets (EPIC-Kitchens-100, SoccerNet). No new human data collection was performed.",
  "rai:dataCollectionRawData": "Synthetic split: no raw data \u2014 streams are sampled from an n-gram generator over a 16-symbol alphabet with the seed-and-resample procedure described in the paper (Appendix 'Benchmark Details'). Natural-video split: raw data are EPIC-Kitchens-100 trimmed action clips (CC BY-NC 4.0, redistributed here) and SoccerNet broadcast events (NDA-gated; only provenance pointers are redistributed, not pixels).",
  "rai:dataCollectionTimeFrame": [
@@ -87,7 +87,7 @@
    "@value": "2026-05-03"
   }
  ],
- "rai:dataCollectionMissingData": "None within the synthetic split (the generator is deterministic given a seed). Natural-video coverage is uneven by design: SoccerNet rows are provided as provenance pointers only (NDA-gated source), so users without SoccerNet access cannot run those rows. The natural-video subset is small (1,028 clips), so per-class statistics may be noisy.",
+ "rai:dataCollectionMissingData": "None within the synthetic split (the generator is deterministic given a seed). Natural-video coverage is uneven by design: SoccerNet rows are provided as provenance pointers only (NDA-gated source), so users without SoccerNet access cannot run those rows. The natural-video subset is small (1,015 clips), so per-class statistics may be noisy.",
  "rai:dataAnnotationProtocol": "No human annotators. Synthetic-stream membership labels (yes/no whether a candidate subsequence appears as a contiguous subsequence of the host stream) are derived programmatically from the known generative process: positives are exact substrings sampled from the host stream and negatives are sampled from the same generator (low/medium bands) or i.i.d. uniform (max-entropy band) and rejection-checked against the host stream. EasyHuman labels are derived from the candidate_tag suffix ('_present' / '_absent'). Natural-video labels are derived from EPIC-Kitchens-100 narration metadata and SoccerNet event annotations as released by the upstream projects; no manual relabeling was performed by this work.",
  "rai:dataAnnotationPlatform": "Programmatic. No annotation platform, crowdsourcing service, or human-in-the-loop tool was used.",
  "rai:dataAnnotationAnalysis": "Annotation correctness is a property of the generator and the upstream metadata, not of human labels: synthetic positive/negative labels are exact (verified via the generator's own substring check at construction time); EasyHuman labels are exact by construction. For the natural-video subset, label quality inherits from the upstream EPIC-Kitchens-100 and SoccerNet annotation pipelines \u2014 no inter-annotator agreement statistics are computed for this work because no new annotation was performed.",
@@ -99,7 +99,7 @@
  "rai:dataSafetyMeasures": "(1) Synthetic streams (text and rendered video) and EasyHuman pose no foreseeable safety risk: content is an abstract 16-symbol alphabet and contains no natural language, no imagery of people, and no operational instructions. (2) Natural-video clips inherit the safety posture of their upstream sources \u2014 EPIC-Kitchens-100 (egocentric cooking, CC BY-NC 4.0, released by the original authors with participant consent) and SoccerNet (broadcast sports footage of public events, distributed under the SoccerNet NDA). No additional automated content moderation, NSFW filtering, or toxicity scanning was applied beyond inheriting these upstream release policies. (3) Membership-query labels are deterministic substring checks over symbol streams and cannot encode harmful content. (4) Users redistributing this dataset, fine-tuning on it, or deploying models evaluated on it should re-run any safety review appropriate to their deployment context; the upstream sources' own safety considerations carry over and are not superseded by this packaging. (5) The dataset ships no model weights and no generated text outputs from evaluated models, so downstream safety risks from model misuse are out of scope of this artifact.",
  "rai:dataUseCases": "Intended uses: (1) evaluating long-context recall in vision-language and language models on a controlled substream-membership task; (2) measuring how recall accuracy varies with input length L and with realized input entropy; (3) studying calibration and abstention behavior under known ground truth; (4) analyzing FLOPs-vs-accuracy trade-offs across model families using the closed-form per-model FLOPs predictor that ships alongside the paper. The benchmark is held out as an evaluation set; it is not intended as training data.",
  "rai:outOfScopeUses": "(1) Not for training data: the benchmark is held out for evaluation, and using it as a training set invalidates its use as a recall benchmark for the trained model. (2) Not for production retrieval, search, or memory systems: synthetic streams are abstract symbols rather than natural language or natural video, and the yes/no substring-membership query format is contrived to isolate recall \u2014 strong scores do not transfer to information-retrieval deployment. (3) Not for general video question-answering evaluation: the natural-video subset is narrow (cooking activities and soccer broadcasts) and the question templates are restricted to membership probes. (4) Not for human-subjects research, biometric identification, face recognition, or surveillance applications: the upstream natural-video sources are not consented for individual identification, and no per-person identifiers are provided. (5) Not for fairness audits or demographic bias measurement: the natural-video subset reflects upstream EPIC-Kitchens-100 and SoccerNet skews (Western kitchens, English-speaking participants, men's professional European football) rather than a controlled demographic sample. (6) Not for safety-critical decision-making about deployed systems: the benchmark targets recall-vs-compute trade-offs, not safety, hallucination calibration outside abstention, or robustness to adversarial inputs. (7) Not for claims about general 'long-term memory' or general video understanding: the controlled, abstract substream task is intentionally narrower than those constructs.",
- "rai:dataLimitations": "(1) The synthetic alphabet is small (16 symbols) and carries no semantic meaning, so strong performance on this benchmark does not directly imply strong performance on natural-language or natural-video memory tasks. (2) The natural-video subset is small (1,028 clips) and skewed toward EPIC-Kitchens-100 cooking activities. (3) SoccerNet rows ship as provenance metadata only; users without an executed SoccerNet NDA cannot run those rows. (4) Compute is reported via a closed-form FLOPs predictor that may diverge from torch.profiler-measured FLOPs at the per-call level (predictor validation is reported in the paper's appendix). (5) Each (length, entropy band) cell at L >= 1024 contains a small number of streams (3-8), so per-cell standard errors at the longest contexts are wide. (6) No human-baseline numbers are reported for the synthetic stream task. (7) The dataset is not designed for safety evaluation, fairness auditing, or production retrieval.",
+ "rai:dataLimitations": "(1) The synthetic alphabet is small (16 symbols) and carries no semantic meaning, so strong performance on this benchmark does not directly imply strong performance on natural-language or natural-video memory tasks. (2) The natural-video subset is small (1,015 clips) and skewed toward EPIC-Kitchens-100 cooking activities. (3) SoccerNet rows ship as provenance metadata only; users without an executed SoccerNet NDA cannot run those rows. (4) Compute is reported via a closed-form FLOPs predictor that may diverge from torch.profiler-measured FLOPs at the per-call level (predictor validation is reported in the paper's appendix). (5) Each (length, entropy band) cell at L >= 1024 contains a small number of streams (3-8), so per-cell standard errors at the longest contexts are wide. (6) No human-baseline numbers are reported for the synthetic stream task. (7) The dataset is not designed for safety evaluation, fairness auditing, or production retrieval.",
  "rai:dataReleaseMaintenancePlan": "Versioned via Hugging Face git commits at https://huggingface.co/datasets/anonstreammem/substream-recollection. The current release is version 2.0.0. Schema or content changes will be recorded in the dataset README and as Hugging Face commit messages on the dataset repo. There is no formal long-term maintenance commitment beyond the paper review and post-acceptance period; dependent upstream sources (EPIC-Kitchens-100, SoccerNet) maintain their own versioning policies.",
  "rai:dataSocialImpact": "This benchmark is intended to support more rigorous evaluation of long-context multimodal models, in particular by separating raw recall from background knowledge and by exposing the FLOPs cost of that recall. The principal foreseeable risk is interpretive: strong scores on this controlled, abstract substream task should not be taken as evidence of general 'long-term memory' or general video understanding. The paper documents concrete failure modes and entropy-dependent accuracy drops to mitigate over-claiming. The dataset is not suitable for fairness audits or for any safety-critical decision-making about deployed systems.",
  "rai:dataBiases": "(1) Synthetic-stream content has no demographic bias by construction (abstract 16-symbol alphabet, uniform sampling within each entropy band), but the choice of three entropy bands and a fixed lane geometry constrains the kinds of memory failure that can be observed. (2) Natural-video biases are inherited from the upstream sources: EPIC-Kitchens-100 over-represents Western kitchens and English-speaking participants and is drawn from a non-representative sample of households; SoccerNet over-represents major European professional leagues and male players, with limited coverage of women's football and non-elite competitions. (3) Question-template coverage is biased toward yes/no membership probes; we do not test free-form recall, multi-step reasoning over the stream, or counterfactual queries.",
@@ -131,8 +131,8 @@
    "description": "Parquet file for the natural_video config (EPIC-Kitchens-100 derived clips and SoccerNet provenance metadata).",
    "contentUrl": "https://huggingface.co/datasets/anonstreammem/substream-recollection/resolve/main/natural_video/questions.parquet",
    "encodingFormat": "application/x-parquet",
-   "contentSize": "162113",
-   "sha256": "616a90fcf2aa18535c59b34cb36e1bf67b6a0f12c5635b45fea35cb7e8064978"
+   "contentSize": "160120",
+   "sha256": "217b2643c7d27a5ea2f31b990dd81024d6741c427bdaea301c55cbbfcc7be760"
   },
   {
    "@type": "cr:FileObject",
@@ -168,11 +168,11 @@
    "@type": "cr:FileObject",
    "@id": "natural_video-manifest-json",
    "name": "natural_video-manifest-json",
-   "description": "Nested-shape manifest.json for the natural_video config \u2014 directly ingestible by the project's main.py. One video per question (1028 entries). No token/lane structure.",
+   "description": "Nested-shape manifest.json for the natural_video config \u2014 directly ingestible by the project's main.py. One video per question (1015 entries). No token/lane structure.",
    "contentUrl": "https://huggingface.co/datasets/anonstreammem/substream-recollection/resolve/main/natural_video/manifest.json",
    "encodingFormat": "application/json",
-   "contentSize": "1566900",
-   "sha256": "afc4dd0912ee9bca4c04ccfb3f1559751f5ac1f5e35625a8fe0dee4992381b89"
+   "contentSize": "1738953",
+   "sha256": "e38e886b10a8a46119e48cab626b941dd9253967659fab780cdee1568f0c82ea"
   },
   {
    "@type": "cr:FileObject",
@@ -877,7 +877,7 @@
    "@type": "cr:RecordSet",
    "@id": "natural_video",
    "name": "natural_video",
-   "description": "Natural-video membership questions over 1028 EPIC-Kitchens-100 and SoccerNet clips at L in {8,16,64,128,512,1024}, 1 fps. Two disjoint question sets, told apart by question_set: 'designed' (1028 rows, the question each clip was built for) and 'density_topup' (785 rows, extra mined questions about other classes, on the 403 codec-banded clips at L in {64,128,512,1024}). No question_id and no (clip, question) pair is shared between them. density_band (low/high) comes from CODEC DISTORTION at a fixed 32 kbit/frame budget (x264, two-pass, one global cutoff on pat_mean held constant at every length); it supersedes an earlier consec_cos_dist measure, which disagreed on 23% of rows and must not be used. density_band is null at L=8 and L=16, buckets that predate that protocol. 166 rows reference NDA-gated SoccerNet clips whose mp4s are not redistributed and are marked media_included=false.",
+   "description": "Natural-video membership questions over 1015 EPIC-Kitchens-100 and SoccerNet clips at L in {8,16,64,128,512,1024}, 1 fps. Two disjoint question sets, told apart by question_set: 'designed' (1007 rows, the question each clip was built for) and 'density_topup' (719 rows, extra mined questions about other classes, on 377 codec-banded clips at L in {64,128,512,1024}). No question_id and no (clip, question) pair is shared between them. density_band (low/high) comes from CODEC DISTORTION at a fixed 32 kbit/frame budget (x264, two-pass, one global cutoff on pat_mean held constant at every length); it supersedes an earlier consec_cos_dist measure, which disagreed on 23% of rows and must not be used. density_band is null at L=8 and L=16, buckets that predate that protocol. 161 rows reference NDA-gated SoccerNet clips whose mp4s are not redistributed and are marked media_included=false.",
    "field": [
     {
      "@type": "cr:Field",
@@ -1349,7 +1349,7 @@
      "@type": "cr:Field",
      "@id": "natural_video/label_source",
      "name": "label_source",
-     "description": "How the label was derived: 'designed' for the designed question set, or narration_margin / whole_video_absence / labels_v2_window / labels_v2_window_absence for the mined ones.",
+     "description": "How the label was derived: 'designed' for the designed question set, or narration_margin / whole_video_absence / labels_v2_window / labels_v2_window_absence for the mined ones; 'annotation_corrected' where the answer was re-checked against the source annotations and corrected.",
      "dataType": "sc:Text",
      "source": {
       "fileObject": {
@@ -1419,6 +1419,51 @@
        "column": "question_set"
       }
      }
+    },
+    {
+     "@type": "cr:Field",
+     "@id": "natural_video/in_density_comparison",
+     "name": "in_density_comparison",
+     "description": "Whether the paper's natural-video density comparison uses this question (the accuracy-versus-compute figure and the band-composition and band-contrast tables). At L=512 and L=1024 the question subsets differ by model; true means at least one model.",
+     "dataType": "sc:Boolean",
+     "source": {
+      "fileObject": {
+       "@id": "natural_video-parquet"
+      },
+      "extract": {
+       "column": "in_density_comparison"
+      }
+     }
+    },
+    {
+     "@type": "cr:Field",
+     "@id": "natural_video/in_calibration",
+     "name": "in_calibration",
+     "description": "Whether the paper's natural-video calibration experiment uses this question (the coverage and selectivity figures and the calibration table).",
+     "dataType": "sc:Boolean",
+     "source": {
+      "fileObject": {
+       "@id": "natural_video-parquet"
+      },
+      "extract": {
+       "column": "in_calibration"
+      }
+     }
+    },
+    {
+     "@type": "cr:Field",
+     "@id": "natural_video/in_rag_comparison",
+     "name": "in_rag_comparison",
+     "description": "Whether the paper's retrieval (RAG) versus full-video comparison uses this question.",
+     "dataType": "sc:Boolean",
+     "source": {
+      "fileObject": {
+       "@id": "natural_video-parquet"
+      },
+      "extract": {
+       "column": "in_rag_comparison"
+      }
+     }
     }
    ]
   },
@@ -1747,5 +1792,5 @@
    ]
   }
  ],
- "dateModified": "2026-08-11"
+ "dateModified": "2026-10-13"
 }

natural_video/questions.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:616a90fcf2aa18535c59b34cb36e1bf67b6a0f12c5635b45fea35cb7e8064978
-size 162090
+oid sha256:217b2643c7d27a5ea2f31b990dd81024d6741c427bdaea301c55cbbfcc7be760
+size 160120