anonstreammem/substream-recollection

Substream Recollection

A controlled benchmark for membership recall, designed to test properties of memory beyond accuracy in LLMs. Each row is a (stream, probe, label) tuple: the model sees a long input stream and a short candidate, and answers whether the candidate (or in the case of natural video, the referenced action) occurred inside the stream.

config	rows	content
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
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

from datasets import load_dataset
ds = load_dataset("anonstreammem/substream-recollection", "natural_video")["train"]

Each config also ships questions.json (NDJSON) and manifest.json (nested, for direct main.py ingestion).

video_path and clip_path are relative to the config that holds the media, given by the media_config column, present on every config:

from huggingface_hub import snapshot_download
root = snapshot_download("anonstreammem/substream-recollection", repo_type="dataset")
path = f"{root}/{row['media_config']}/{row['video_path']}"     # works for every config

Check media_included first: false means the mp4 is NDA-gated and absent, null means the row has no media. Filter with media_included != False, not df[df.media_included], or the null rows are dropped too.

Columns worth knowing

answer is the ground truth, "yes" / "no".

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 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)

They are disjoint: no shared question_id, and no clip is asked the same question twice. Filter on question_set to recover either subset.

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
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_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
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

Event columns localise the clip's designed event, which is what this question asks about only when question_set == "designed". No density_topup question asks about its clip's designed class, so for those rows compare event_type with query_class before reading the event columns as the answer's location.

89 groups of clips are byte-identical under different names — a clip built as one class's positive is reused as another's negative. Group on clip_sha256 to dedupe or check independence; contains_anomaly and the event columns describe the named clip, not the file.

Licences

Synthetic and EasyHuman content is CC BY 4.0. EPIC-Kitchens-100 derived clips are CC BY-NC 4.0. SoccerNet rows are provenance-only under the SoccerNet NDA. Per-row license column; full texts in LICENSES/.

Citation

Anonymous, "Good Memory Has ECC: Evaluating the Memory of Vision-Language Models Beyond Accuracy," 2026 (anonymized for review).

Cite the original sources for the natural videos:

• Damen, D. et al. "Rescaling Egocentric Vision: Collection, Pipeline and Challenges for EPIC-KITCHENS-100." IJCV, 2022.
• Giancola, S., Amine, M., Dghaily, T., Ghanem, B. "SoccerNet: A Scalable Dataset for Action Spotting in Soccer Videos." CVPR Workshops, 2018.